diff --git a/iopaint/model/base.py b/iopaint/model/base.py
index 8d79179..433ad68 100644
--- a/iopaint/model/base.py
+++ b/iopaint/model/base.py
@@ -13,7 +13,7 @@ from iopaint.helper import (
switch_mps_device,
)
from iopaint.schema import InpaintRequest, HDStrategy, SDSampler
-from .helper.g_diffuser_bot import expand_image, expand_image2
+from .helper.g_diffuser_bot import expand_image
from .utils import get_scheduler
@@ -35,8 +35,7 @@ class InpaintModel:
self.init_model(device, **kwargs)
@abc.abstractmethod
- def init_model(self, device, **kwargs):
- ...
+ def init_model(self, device, **kwargs): ...
@staticmethod
@abc.abstractmethod
@@ -53,8 +52,7 @@ class InpaintModel:
...
@staticmethod
- def download():
- ...
+ def download(): ...
def _pad_forward(self, image, mask, config: InpaintRequest):
origin_height, origin_width = image.shape[:2]
@@ -96,7 +94,7 @@ class InpaintModel:
# logger.info(f"hd_strategy: {config.hd_strategy}")
if config.hd_strategy == HDStrategy.CROP:
if max(image.shape) > config.hd_strategy_crop_trigger_size:
- logger.info(f"Run crop strategy")
+ logger.info("Run crop strategy")
boxes = boxes_from_mask(mask)
crop_result = []
for box in boxes:
@@ -327,14 +325,12 @@ class DiffusionInpaintModel(InpaintModel):
padding_r = max(0, cropper_r - image_r)
padding_b = max(0, cropper_b - image_b)
- expanded_image, mask_image = expand_image2(
+ expanded_image, mask_image = expand_image(
cropped_image,
left=padding_l,
top=padding_t,
right=padding_r,
bottom=padding_b,
- softness=config.sd_outpainting_softness,
- space=config.sd_outpainting_space,
)
# 最终扩大了的 image, BGR
@@ -381,15 +377,6 @@ class DiffusionInpaintModel(InpaintModel):
interpolation=cv2.INTER_CUBIC,
)
- # blend result, copy from g_diffuser_bot
- # mask_rgb = 1.0 - np_img_grey_to_rgb(mask / 255.0)
- # inpaint_result = np.clip(
- # inpaint_result * (1.0 - mask_rgb) + image * mask_rgb, 0.0, 255.0
- # )
- # original_pixel_indices = mask < 127
- # inpaint_result[original_pixel_indices] = image[:, :, ::-1][
- # original_pixel_indices
- # ]
return inpaint_result
def set_scheduler(self, config: InpaintRequest):
@@ -412,7 +399,7 @@ class DiffusionInpaintModel(InpaintModel):
if config.sd_match_histograms:
result = self._match_histograms(result, image[:, :, ::-1], mask)
- # if config.sd_mask_blur != 0:
- # k = 2 * config.sd_mask_blur + 1
- # mask = cv2.GaussianBlur(mask, (k, k), 0)
+ if config.use_extender and config.sd_mask_blur != 0:
+ k = 2 * config.sd_mask_blur + 1
+ mask = cv2.GaussianBlur(mask, (k, k), 0)
return result, image, mask
diff --git a/iopaint/model/helper/g_diffuser_bot.py b/iopaint/model/helper/g_diffuser_bot.py
index f669f9a..79b19aa 100644
--- a/iopaint/model/helper/g_diffuser_bot.py
+++ b/iopaint/model/helper/g_diffuser_bot.py
@@ -1,174 +1,29 @@
-# code copy from: https://github.com/parlance-zz/g-diffuser-bot
import cv2
import numpy as np
-def np_img_grey_to_rgb(data):
- if data.ndim == 3:
- return data
- return np.expand_dims(data, 2) * np.ones((1, 1, 3))
-
-
-def convolve(data1, data2): # fast convolution with fft
- if data1.ndim != data2.ndim: # promote to rgb if mismatch
- if data1.ndim < 3:
- data1 = np_img_grey_to_rgb(data1)
- if data2.ndim < 3:
- data2 = np_img_grey_to_rgb(data2)
- return ifft2(fft2(data1) * fft2(data2))
-
-
-def fft2(data):
- if data.ndim > 2: # multiple channels
- out_fft = np.zeros(
- (data.shape[0], data.shape[1], data.shape[2]), dtype=np.complex128
- )
- for c in range(data.shape[2]):
- c_data = data[:, :, c]
- out_fft[:, :, c] = np.fft.fft2(np.fft.fftshift(c_data), norm="ortho")
- out_fft[:, :, c] = np.fft.ifftshift(out_fft[:, :, c])
- else: # single channel
- out_fft = np.zeros((data.shape[0], data.shape[1]), dtype=np.complex128)
- out_fft[:, :] = np.fft.fft2(np.fft.fftshift(data), norm="ortho")
- out_fft[:, :] = np.fft.ifftshift(out_fft[:, :])
-
- return out_fft
-
-
-def ifft2(data):
- if data.ndim > 2: # multiple channels
- out_ifft = np.zeros(
- (data.shape[0], data.shape[1], data.shape[2]), dtype=np.complex128
- )
- for c in range(data.shape[2]):
- c_data = data[:, :, c]
- out_ifft[:, :, c] = np.fft.ifft2(np.fft.fftshift(c_data), norm="ortho")
- out_ifft[:, :, c] = np.fft.ifftshift(out_ifft[:, :, c])
- else: # single channel
- out_ifft = np.zeros((data.shape[0], data.shape[1]), dtype=np.complex128)
- out_ifft[:, :] = np.fft.ifft2(np.fft.fftshift(data), norm="ortho")
- out_ifft[:, :] = np.fft.ifftshift(out_ifft[:, :])
-
- return out_ifft
-
-
-def get_gradient_kernel(width, height, std=3.14, mode="linear"):
- window_scale_x = float(
- width / min(width, height)
- ) # for non-square aspect ratios we still want a circular kernel
- window_scale_y = float(height / min(width, height))
- if mode == "gaussian":
- x = (np.arange(width) / width * 2.0 - 1.0) * window_scale_x
- kx = np.exp(-x * x * std)
- if window_scale_x != window_scale_y:
- y = (np.arange(height) / height * 2.0 - 1.0) * window_scale_y
- ky = np.exp(-y * y * std)
- else:
- y = x
- ky = kx
- return np.outer(kx, ky)
- elif mode == "linear":
- x = (np.arange(width) / width * 2.0 - 1.0) * window_scale_x
- if window_scale_x != window_scale_y:
- y = (np.arange(height) / height * 2.0 - 1.0) * window_scale_y
- else:
- y = x
- return np.clip(1.0 - np.sqrt(np.add.outer(x * x, y * y)) * std / 3.14, 0.0, 1.0)
- else:
- raise Exception("Error: Unknown mode in get_gradient_kernel: {0}".format(mode))
-
-
-def image_blur(data, std=3.14, mode="linear"):
- width = data.shape[0]
- height = data.shape[1]
- kernel = get_gradient_kernel(width, height, std, mode=mode)
- return np.real(convolve(data, kernel / np.sqrt(np.sum(kernel * kernel))))
-
-
-def soften_mask(mask_img, softness, space):
- if softness == 0:
- return mask_img
- softness = min(softness, 1.0)
- space = np.clip(space, 0.0, 1.0)
- original_max_opacity = np.max(mask_img)
- out_mask = mask_img <= 0.0
- blurred_mask = image_blur(mask_img, 3.5 / softness, mode="linear")
- blurred_mask = np.maximum(blurred_mask - np.max(blurred_mask[out_mask]), 0.0)
- mask_img *= blurred_mask # preserve partial opacity in original input mask
- mask_img /= np.max(mask_img) # renormalize
- mask_img = np.clip(mask_img - space, 0.0, 1.0) # make space
- mask_img /= np.max(mask_img) # and renormalize again
- mask_img *= original_max_opacity # restore original max opacity
- return mask_img
-
-
-def expand_image(
- cv2_img, top: int, right: int, bottom: int, left: int, softness: float, space: float
-):
+def expand_image(cv2_img, top: int, right: int, bottom: int, left: int):
assert cv2_img.shape[2] == 3
origin_h, origin_w = cv2_img.shape[:2]
- new_width = cv2_img.shape[1] + left + right
- new_height = cv2_img.shape[0] + top + bottom
# TODO: which is better?
- # new_img = np.random.randint(0, 255, (new_height, new_width, 3), np.uint8)
- new_img = cv2.copyMakeBorder(
- cv2_img, top, bottom, left, right, cv2.BORDER_REPLICATE
- )
- mask_img = np.zeros((new_height, new_width), np.uint8)
- mask_img[top: top + cv2_img.shape[0], left: left + cv2_img.shape[1]] = 255
-
- if softness > 0.0:
- mask_img = soften_mask(mask_img / 255.0, softness / 100.0, space / 100.0)
- mask_img = (np.clip(mask_img, 0.0, 1.0) * 255.0).astype(np.uint8)
-
- mask_image = 255.0 - mask_img # extract mask from alpha channel and invert
- rgb_init_image = (
- 0.0 + new_img[:, :, 0:3]
- ) # strip mask from init_img leaving only rgb channels
-
- hard_mask = np.zeros_like(cv2_img[:, :, 0])
- if top != 0:
- hard_mask[0: origin_h // 2, :] = 255
- if bottom != 0:
- hard_mask[origin_h // 2:, :] = 255
- if left != 0:
- hard_mask[:, 0: origin_w // 2] = 255
- if right != 0:
- hard_mask[:, origin_w // 2:] = 255
-
- hard_mask = cv2.copyMakeBorder(
- hard_mask, top, bottom, left, right, cv2.BORDER_CONSTANT, value=255
- )
- mask_image = np.where(hard_mask > 0, mask_image, 0)
- return rgb_init_image.astype(np.uint8), mask_image.astype(np.uint8)
-
-
-def expand_image2(
- cv2_img, top: int, right: int, bottom: int, left: int, softness: float, space: float
-):
- assert cv2_img.shape[2] == 3
- origin_h, origin_w = cv2_img.shape[:2]
- new_width = cv2_img.shape[1] + left + right
- new_height = cv2_img.shape[0] + top + bottom
-
- # TODO: which is better?
- # new_img = np.random.randint(0, 255, (new_height, new_width, 3), np.uint8)
+ # new_img = np.ones((new_height, new_width, 3), np.uint8) * 255
new_img = cv2.copyMakeBorder(
cv2_img, top, bottom, left, right, cv2.BORDER_REPLICATE
)
- inner_padding_left = 13 if left > 0 else 0
- inner_padding_right = 13 if right > 0 else 0
- inner_padding_top = 13 if top > 0 else 0
- inner_padding_bottom = 13 if bottom > 0 else 0
+ inner_padding_left = 0 if left > 0 else 0
+ inner_padding_right = 0 if right > 0 else 0
+ inner_padding_top = 0 if top > 0 else 0
+ inner_padding_bottom = 0 if bottom > 0 else 0
mask_image = np.zeros(
(
- origin_h - inner_padding_top - inner_padding_bottom
- , origin_w - inner_padding_left - inner_padding_right
+ origin_h - inner_padding_top - inner_padding_bottom,
+ origin_w - inner_padding_left - inner_padding_right,
),
- np.uint8)
+ np.uint8,
+ )
mask_image = cv2.copyMakeBorder(
mask_image,
top + inner_padding_top,
@@ -176,11 +31,11 @@ def expand_image2(
left + inner_padding_left,
right + inner_padding_right,
cv2.BORDER_CONSTANT,
- value=255
+ value=255,
)
# k = 2*int(min(origin_h, origin_w) // 6)+1
- k = 7
- mask_image = cv2.GaussianBlur(mask_image, (k, k), 0)
+ # k = 7
+ # mask_image = cv2.GaussianBlur(mask_image, (k, k), 0)
return new_img, mask_image
@@ -190,7 +45,7 @@ if __name__ == "__main__":
current_dir = Path(__file__).parent.absolute().resolve()
image_path = "/Users/cwq/code/github/IOPaint/iopaint/tests/bunny.jpeg"
init_image = cv2.imread(str(image_path))
- init_image, mask_image = expand_image2(
+ init_image, mask_image = expand_image(
init_image,
top=0,
right=0,
diff --git a/iopaint/model/power_paint/pipeline_powerpaint_controlnet.py b/iopaint/model/power_paint/pipeline_powerpaint_controlnet.py
deleted file mode 100644
index cba0f8f..0000000
--- a/iopaint/model/power_paint/pipeline_powerpaint_controlnet.py
+++ /dev/null
@@ -1,1775 +0,0 @@
-# Copyright 2023 The HuggingFace Team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-# http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-# This model implementation is heavily inspired by https://github.com/haofanwang/ControlNet-for-Diffusers/
-
-import inspect
-import warnings
-from typing import Any, Callable, Dict, List, Optional, Tuple, Union
-
-import numpy as np
-import PIL.Image
-import torch
-import torch.nn.functional as F
-from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
-
-from diffusers.image_processor import VaeImageProcessor
-from diffusers.loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin
-from diffusers.models import AutoencoderKL, ControlNetModel, UNet2DConditionModel
-from diffusers.schedulers import KarrasDiffusionSchedulers
-from diffusers.utils import (
- is_accelerate_available,
- is_accelerate_version,
- logging,
- replace_example_docstring,
-)
-from diffusers.utils.torch_utils import randn_tensor,is_compiled_module
-from diffusers.pipelines.pipeline_utils import DiffusionPipeline
-from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
-from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
-from diffusers.pipelines.controlnet import MultiControlNetModel
-
-
-
-
-logger = logging.get_logger(__name__) # pylint: disable=invalid-name
-
-
-EXAMPLE_DOC_STRING = """
- Examples:
- ```py
- >>> # !pip install transformers accelerate
- >>> from diffusers import StableDiffusionControlNetInpaintPipeline, ControlNetModel, DDIMScheduler
- >>> from diffusers.utils import load_image
- >>> import numpy as np
- >>> import torch
-
- >>> init_image = load_image(
- ... "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main/stable_diffusion_inpaint/boy.png"
- ... )
- >>> init_image = init_image.resize((512, 512))
-
- >>> generator = torch.Generator(device="cpu").manual_seed(1)
-
- >>> mask_image = load_image(
- ... "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main/stable_diffusion_inpaint/boy_mask.png"
- ... )
- >>> mask_image = mask_image.resize((512, 512))
-
-
- >>> def make_inpaint_condition(image, image_mask):
- ... image = np.array(image.convert("RGB")).astype(np.float32) / 255.0
- ... image_mask = np.array(image_mask.convert("L")).astype(np.float32) / 255.0
-
- ... assert image.shape[0:1] == image_mask.shape[0:1], "image and image_mask must have the same image size"
- ... image[image_mask > 0.5] = -1.0 # set as masked pixel
- ... image = np.expand_dims(image, 0).transpose(0, 3, 1, 2)
- ... image = torch.from_numpy(image)
- ... return image
-
-
- >>> control_image = make_inpaint_condition(init_image, mask_image)
-
- >>> controlnet = ControlNetModel.from_pretrained(
- ... "lllyasviel/control_v11p_sd15_inpaint", torch_dtype=torch.float16
- ... )
- >>> pipe = StableDiffusionControlNetInpaintPipeline.from_pretrained(
- ... "runwayml/stable-diffusion-v1-5", controlnet=controlnet, torch_dtype=torch.float16
- ... )
-
- >>> pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
- >>> pipe.enable_model_cpu_offload()
-
- >>> # generate image
- >>> image = pipe(
- ... "a handsome man with ray-ban sunglasses",
- ... num_inference_steps=20,
- ... generator=generator,
- ... eta=1.0,
- ... image=init_image,
- ... mask_image=mask_image,
- ... control_image=control_image,
- ... ).images[0]
- ```
-"""
-
-
-# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint.prepare_mask_and_masked_image
-def prepare_mask_and_masked_image(image, mask, height, width, return_image=False):
- """
- Prepares a pair (image, mask) to be consumed by the Stable Diffusion pipeline. This means that those inputs will be
- converted to ``torch.Tensor`` with shapes ``batch x channels x height x width`` where ``channels`` is ``3`` for the
- ``image`` and ``1`` for the ``mask``.
-
- The ``image`` will be converted to ``torch.float32`` and normalized to be in ``[-1, 1]``. The ``mask`` will be
- binarized (``mask > 0.5``) and cast to ``torch.float32`` too.
-
- Args:
- image (Union[np.array, PIL.Image, torch.Tensor]): The image to inpaint.
- It can be a ``PIL.Image``, or a ``height x width x 3`` ``np.array`` or a ``channels x height x width``
- ``torch.Tensor`` or a ``batch x channels x height x width`` ``torch.Tensor``.
- mask (_type_): The mask to apply to the image, i.e. regions to inpaint.
- It can be a ``PIL.Image``, or a ``height x width`` ``np.array`` or a ``1 x height x width``
- ``torch.Tensor`` or a ``batch x 1 x height x width`` ``torch.Tensor``.
-
-
- Raises:
- ValueError: ``torch.Tensor`` images should be in the ``[-1, 1]`` range. ValueError: ``torch.Tensor`` mask
- should be in the ``[0, 1]`` range. ValueError: ``mask`` and ``image`` should have the same spatial dimensions.
- TypeError: ``mask`` is a ``torch.Tensor`` but ``image`` is not
- (ot the other way around).
-
- Returns:
- tuple[torch.Tensor]: The pair (mask, masked_image) as ``torch.Tensor`` with 4
- dimensions: ``batch x channels x height x width``.
- """
-
- if image is None:
- raise ValueError("`image` input cannot be undefined.")
-
- if mask is None:
- raise ValueError("`mask_image` input cannot be undefined.")
-
- if isinstance(image, torch.Tensor):
- if not isinstance(mask, torch.Tensor):
- raise TypeError(f"`image` is a torch.Tensor but `mask` (type: {type(mask)} is not")
-
- # Batch single image
- if image.ndim == 3:
- assert image.shape[0] == 3, "Image outside a batch should be of shape (3, H, W)"
- image = image.unsqueeze(0)
-
- # Batch and add channel dim for single mask
- if mask.ndim == 2:
- mask = mask.unsqueeze(0).unsqueeze(0)
-
- # Batch single mask or add channel dim
- if mask.ndim == 3:
- # Single batched mask, no channel dim or single mask not batched but channel dim
- if mask.shape[0] == 1:
- mask = mask.unsqueeze(0)
-
- # Batched masks no channel dim
- else:
- mask = mask.unsqueeze(1)
-
- assert image.ndim == 4 and mask.ndim == 4, "Image and Mask must have 4 dimensions"
- assert image.shape[-2:] == mask.shape[-2:], "Image and Mask must have the same spatial dimensions"
- assert image.shape[0] == mask.shape[0], "Image and Mask must have the same batch size"
-
- # Check image is in [-1, 1]
- if image.min() < -1 or image.max() > 1:
- raise ValueError("Image should be in [-1, 1] range")
-
- # Check mask is in [0, 1]
- if mask.min() < 0 or mask.max() > 1:
- raise ValueError("Mask should be in [0, 1] range")
-
- # Binarize mask
- mask[mask < 0.5] = 0
- mask[mask >= 0.5] = 1
-
- # Image as float32
- image = image.to(dtype=torch.float32)
- elif isinstance(mask, torch.Tensor):
- raise TypeError(f"`mask` is a torch.Tensor but `image` (type: {type(image)} is not")
- else:
- # preprocess image
- if isinstance(image, (PIL.Image.Image, np.ndarray)):
- image = [image]
- if isinstance(image, list) and isinstance(image[0], PIL.Image.Image):
- # resize all images w.r.t passed height an width
- image = [i.resize((width, height), resample=PIL.Image.LANCZOS) for i in image]
- image = [np.array(i.convert("RGB"))[None, :] for i in image]
- image = np.concatenate(image, axis=0)
- elif isinstance(image, list) and isinstance(image[0], np.ndarray):
- image = np.concatenate([i[None, :] for i in image], axis=0)
-
- image = image.transpose(0, 3, 1, 2)
- image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0
-
- # preprocess mask
- if isinstance(mask, (PIL.Image.Image, np.ndarray)):
- mask = [mask]
-
- if isinstance(mask, list) and isinstance(mask[0], PIL.Image.Image):
- mask = [i.resize((width, height), resample=PIL.Image.LANCZOS) for i in mask]
- mask = np.concatenate([np.array(m.convert("L"))[None, None, :] for m in mask], axis=0)
- mask = mask.astype(np.float32) / 255.0
- elif isinstance(mask, list) and isinstance(mask[0], np.ndarray):
- mask = np.concatenate([m[None, None, :] for m in mask], axis=0)
-
- mask[mask < 0.5] = 0
- mask[mask >= 0.5] = 1
- mask = torch.from_numpy(mask)
-
- masked_image = image * (mask < 0.5)
-
- # n.b. ensure backwards compatibility as old function does not return image
- if return_image:
- return mask, masked_image, image
-
- return mask, masked_image
-
-
-class StableDiffusionControlNetInpaintPipeline(
- DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, FromSingleFileMixin
-):
- r"""
- Pipeline for text-to-image generation using Stable Diffusion with ControlNet guidance.
-
- This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
- library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
-
- In addition the pipeline inherits the following loading methods:
- - *Textual-Inversion*: [`loaders.TextualInversionLoaderMixin.load_textual_inversion`]
-
-
-
- This pipeline can be used both with checkpoints that have been specifically fine-tuned for inpainting, such as
- [runwayml/stable-diffusion-inpainting](https://huggingface.co/runwayml/stable-diffusion-inpainting)
- as well as default text-to-image stable diffusion checkpoints, such as
- [runwayml/stable-diffusion-v1-5](https://huggingface.co/runwayml/stable-diffusion-v1-5).
- Default text-to-image stable diffusion checkpoints might be preferable for controlnets that have been fine-tuned on
- those, such as [lllyasviel/control_v11p_sd15_inpaint](https://huggingface.co/lllyasviel/control_v11p_sd15_inpaint).
-
-
-
- Args:
- vae ([`AutoencoderKL`]):
- Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
- text_encoder ([`CLIPTextModel`]):
- Frozen text-encoder. Stable Diffusion uses the text portion of
- [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
- the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
- tokenizer (`CLIPTokenizer`):
- Tokenizer of class
- [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
- unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
- controlnet ([`ControlNetModel`] or `List[ControlNetModel]`):
- Provides additional conditioning to the unet during the denoising process. If you set multiple ControlNets
- as a list, the outputs from each ControlNet are added together to create one combined additional
- conditioning.
- scheduler ([`SchedulerMixin`]):
- A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
- [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
- safety_checker ([`StableDiffusionSafetyChecker`]):
- Classification module that estimates whether generated images could be considered offensive or harmful.
- Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
- feature_extractor ([`CLIPImageProcessor`]):
- Model that extracts features from generated images to be used as inputs for the `safety_checker`.
- """
- _optional_components = ["safety_checker", "feature_extractor"]
-
- def __init__(
- self,
- vae: AutoencoderKL,
- text_encoder: CLIPTextModel,
- tokenizer: CLIPTokenizer,
- unet: UNet2DConditionModel,
- controlnet: Union[ControlNetModel, List[ControlNetModel], Tuple[ControlNetModel], MultiControlNetModel],
- scheduler: KarrasDiffusionSchedulers,
- safety_checker: StableDiffusionSafetyChecker,
- feature_extractor: CLIPImageProcessor,
- requires_safety_checker: bool = True,
- ):
- super().__init__()
-
- if safety_checker is None and requires_safety_checker:
- logger.warning(
- f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
- " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
- " results in services or applications open to the public. Both the diffusers team and Hugging Face"
- " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
- " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
- " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
- )
-
- if safety_checker is not None and feature_extractor is None:
- raise ValueError(
- "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
- " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
- )
-
- if isinstance(controlnet, (list, tuple)):
- controlnet = MultiControlNetModel(controlnet)
-
- self.register_modules(
- vae=vae,
- text_encoder=text_encoder,
- tokenizer=tokenizer,
- unet=unet,
- controlnet=controlnet,
- scheduler=scheduler,
- safety_checker=safety_checker,
- feature_extractor=feature_extractor,
- )
- self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
- self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
- self.control_image_processor = VaeImageProcessor(
- vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True, do_normalize=False
- )
- self.register_to_config(requires_safety_checker=requires_safety_checker)
-
- # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
- def enable_vae_slicing(self):
- r"""
- Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
- compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
- """
- self.vae.enable_slicing()
-
- # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
- def disable_vae_slicing(self):
- r"""
- Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
- computing decoding in one step.
- """
- self.vae.disable_slicing()
-
- # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
- def enable_vae_tiling(self):
- r"""
- Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
- compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
- processing larger images.
- """
- self.vae.enable_tiling()
-
- # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
- def disable_vae_tiling(self):
- r"""
- Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
- computing decoding in one step.
- """
- self.vae.disable_tiling()
-
- def enable_model_cpu_offload(self, gpu_id=0):
- r"""
- Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
- to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
- method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
- `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
- """
- if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
- from accelerate import cpu_offload_with_hook
- else:
- raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.")
-
- device = torch.device(f"cuda:{gpu_id}")
-
- hook = None
- for cpu_offloaded_model in [self.text_encoder, self.unet, self.vae]:
- _, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook)
-
- if self.safety_checker is not None:
- # the safety checker can offload the vae again
- _, hook = cpu_offload_with_hook(self.safety_checker, device, prev_module_hook=hook)
-
- # control net hook has be manually offloaded as it alternates with unet
- cpu_offload_with_hook(self.controlnet, device)
-
- # We'll offload the last model manually.
- self.final_offload_hook = hook
-
- # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
- def _encode_prompt(
- self,
- promptA,
- promptB,
- t,
- device,
- num_images_per_prompt,
- do_classifier_free_guidance,
- negative_promptA=None,
- negative_promptB=None,
- t_nag = None,
- prompt_embeds: Optional[torch.FloatTensor] = None,
- negative_prompt_embeds: Optional[torch.FloatTensor] = None,
- lora_scale: Optional[float] = None,
- ):
- r"""
- Encodes the prompt into text encoder hidden states.
-
- Args:
- prompt (`str` or `List[str]`, *optional*):
- prompt to be encoded
- device: (`torch.device`):
- torch device
- num_images_per_prompt (`int`):
- number of images that should be generated per prompt
- do_classifier_free_guidance (`bool`):
- whether to use classifier free guidance or not
- negative_prompt (`str` or `List[str]`, *optional*):
- The prompt or prompts not to guide the image generation. If not defined, one has to pass
- `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
- less than `1`).
- prompt_embeds (`torch.FloatTensor`, *optional*):
- Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
- provided, text embeddings will be generated from `prompt` input argument.
- negative_prompt_embeds (`torch.FloatTensor`, *optional*):
- Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
- weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
- argument.
- lora_scale (`float`, *optional*):
- A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
- """
- # set lora scale so that monkey patched LoRA
- # function of text encoder can correctly access it
- if lora_scale is not None and isinstance(self, LoraLoaderMixin):
- self._lora_scale = lora_scale
-
- prompt = promptA
- negative_prompt = negative_promptA
-
- if promptA is not None and isinstance(promptA, str):
- batch_size = 1
- elif promptA is not None and isinstance(promptA, list):
- batch_size = len(promptA)
- else:
- batch_size = prompt_embeds.shape[0]
-
- if prompt_embeds is None:
- # textual inversion: procecss multi-vector tokens if necessary
- if isinstance(self, TextualInversionLoaderMixin):
- promptA = self.maybe_convert_prompt(promptA, self.tokenizer)
-
- text_inputsA = self.tokenizer(
- promptA,
- padding="max_length",
- max_length=self.tokenizer.model_max_length,
- truncation=True,
- return_tensors="pt",
- )
- text_inputsB = self.tokenizer(
- promptB,
- padding="max_length",
- max_length=self.tokenizer.model_max_length,
- truncation=True,
- return_tensors="pt",
- )
- text_input_idsA = text_inputsA.input_ids
- text_input_idsB = text_inputsB.input_ids
- untruncated_ids = self.tokenizer(promptA, padding="longest", return_tensors="pt").input_ids
-
- if untruncated_ids.shape[-1] >= text_input_idsA.shape[-1] and not torch.equal(
- text_input_idsA, untruncated_ids
- ):
- removed_text = self.tokenizer.batch_decode(
- untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
- )
- logger.warning(
- "The following part of your input was truncated because CLIP can only handle sequences up to"
- f" {self.tokenizer.model_max_length} tokens: {removed_text}"
- )
-
- if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
- attention_mask = text_inputsA.attention_mask.to(device)
- else:
- attention_mask = None
-
- # print("text_input_idsA: ",text_input_idsA)
- # print("text_input_idsB: ",text_input_idsB)
- # print('t: ',t)
-
- prompt_embedsA = self.text_encoder(
- text_input_idsA.to(device),
- attention_mask=attention_mask,
- )
- prompt_embedsA = prompt_embedsA[0]
-
- prompt_embedsB = self.text_encoder(
- text_input_idsB.to(device),
- attention_mask=attention_mask,
- )
- prompt_embedsB = prompt_embedsB[0]
- prompt_embeds = prompt_embedsA*(t)+(1-t)*prompt_embedsB
- # print("prompt_embeds: ",prompt_embeds)
-
- if self.text_encoder is not None:
- prompt_embeds_dtype = self.text_encoder.dtype
- elif self.unet is not None:
- prompt_embeds_dtype = self.unet.dtype
- else:
- prompt_embeds_dtype = prompt_embeds.dtype
-
- prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
-
- bs_embed, seq_len, _ = prompt_embeds.shape
- # duplicate text embeddings for each generation per prompt, using mps friendly method
- prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
- prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
-
- # get unconditional embeddings for classifier free guidance
- if do_classifier_free_guidance and negative_prompt_embeds is None:
- uncond_tokensA: List[str]
- uncond_tokensB: List[str]
- if negative_prompt is None:
- uncond_tokensA = [""] * batch_size
- uncond_tokensB = [""] * batch_size
- elif prompt is not None and type(prompt) is not type(negative_prompt):
- raise TypeError(
- f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
- f" {type(prompt)}."
- )
- elif isinstance(negative_prompt, str):
- uncond_tokensA = [negative_promptA]
- uncond_tokensB = [negative_promptB]
- elif batch_size != len(negative_prompt):
- raise ValueError(
- f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
- f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
- " the batch size of `prompt`."
- )
- else:
- uncond_tokensA = negative_promptA
- uncond_tokensB = negative_promptB
-
- # textual inversion: procecss multi-vector tokens if necessary
- if isinstance(self, TextualInversionLoaderMixin):
- uncond_tokensA = self.maybe_convert_prompt(uncond_tokensA, self.tokenizer)
- uncond_tokensB = self.maybe_convert_prompt(uncond_tokensB, self.tokenizer)
-
- max_length = prompt_embeds.shape[1]
- uncond_inputA = self.tokenizer(
- uncond_tokensA,
- padding="max_length",
- max_length=max_length,
- truncation=True,
- return_tensors="pt",
- )
- uncond_inputB = self.tokenizer(
- uncond_tokensB,
- padding="max_length",
- max_length=max_length,
- truncation=True,
- return_tensors="pt",
- )
-
- if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
- attention_mask = uncond_inputA.attention_mask.to(device)
- else:
- attention_mask = None
-
- negative_prompt_embedsA = self.text_encoder(
- uncond_inputA.input_ids.to(device),
- attention_mask=attention_mask,
- )
- negative_prompt_embedsB = self.text_encoder(
- uncond_inputB.input_ids.to(device),
- attention_mask=attention_mask,
- )
- negative_prompt_embeds = negative_prompt_embedsA[0]*(t_nag)+(1-t_nag)*negative_prompt_embedsB[0]
-
- # negative_prompt_embeds = negative_prompt_embeds[0]
-
- if do_classifier_free_guidance:
- # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
- seq_len = negative_prompt_embeds.shape[1]
-
- negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
-
- negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
- negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
-
- # For classifier free guidance, we need to do two forward passes.
- # Here we concatenate the unconditional and text embeddings into a single batch
- # to avoid doing two forward passes
- # print("prompt_embeds: ",prompt_embeds)
- prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
-
- return prompt_embeds
-
- # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
- def run_safety_checker(self, image, device, dtype):
- if self.safety_checker is None:
- has_nsfw_concept = None
- else:
- if torch.is_tensor(image):
- feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
- else:
- feature_extractor_input = self.image_processor.numpy_to_pil(image)
- safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
- image, has_nsfw_concept = self.safety_checker(
- images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
- )
- return image, has_nsfw_concept
-
- # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
- def decode_latents(self, latents):
- warnings.warn(
- "The decode_latents method is deprecated and will be removed in a future version. Please"
- " use VaeImageProcessor instead",
- FutureWarning,
- )
- latents = 1 / self.vae.config.scaling_factor * latents
- image = self.vae.decode(latents, return_dict=False)[0]
- image = (image / 2 + 0.5).clamp(0, 1)
- # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
- image = image.cpu().permute(0, 2, 3, 1).float().numpy()
- return image
-
- # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
- def prepare_extra_step_kwargs(self, generator, eta):
- # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
- # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
- # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
- # and should be between [0, 1]
-
- accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
- extra_step_kwargs = {}
- if accepts_eta:
- extra_step_kwargs["eta"] = eta
-
- # check if the scheduler accepts generator
- accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
- if accepts_generator:
- extra_step_kwargs["generator"] = generator
- return extra_step_kwargs
-
- # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps
- def get_timesteps(self, num_inference_steps, strength, device):
- # get the original timestep using init_timestep
- init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
-
- t_start = max(num_inference_steps - init_timestep, 0)
- timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
-
- return timesteps, num_inference_steps - t_start
-
- def check_inputs(
- self,
- prompt,
- image,
- height,
- width,
- callback_steps,
- negative_prompt=None,
- prompt_embeds=None,
- negative_prompt_embeds=None,
- controlnet_conditioning_scale=1.0,
- control_guidance_start=0.0,
- control_guidance_end=1.0,
- ):
- if height % 8 != 0 or width % 8 != 0:
- raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
-
- if (callback_steps is None) or (
- callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
- ):
- raise ValueError(
- f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
- f" {type(callback_steps)}."
- )
-
- if prompt is not None and prompt_embeds is not None:
- raise ValueError(
- f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
- " only forward one of the two."
- )
- elif prompt is None and prompt_embeds is None:
- raise ValueError(
- "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
- )
- elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
- raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
-
- if negative_prompt is not None and negative_prompt_embeds is not None:
- raise ValueError(
- f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
- f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
- )
-
- if prompt_embeds is not None and negative_prompt_embeds is not None:
- if prompt_embeds.shape != negative_prompt_embeds.shape:
- raise ValueError(
- "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
- f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
- f" {negative_prompt_embeds.shape}."
- )
-
- # `prompt` needs more sophisticated handling when there are multiple
- # conditionings.
- if isinstance(self.controlnet, MultiControlNetModel):
- if isinstance(prompt, list):
- logger.warning(
- f"You have {len(self.controlnet.nets)} ControlNets and you have passed {len(prompt)}"
- " prompts. The conditionings will be fixed across the prompts."
- )
-
- # Check `image`
- is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance(
- self.controlnet, torch._dynamo.eval_frame.OptimizedModule
- )
-
- if (
- isinstance(self.controlnet, ControlNetModel)
- or is_compiled
- and isinstance(self.controlnet._orig_mod, ControlNetModel)
- ):
- self.check_image(image, prompt, prompt_embeds)
- elif (
- isinstance(self.controlnet, MultiControlNetModel)
- or is_compiled
- and isinstance(self.controlnet._orig_mod, MultiControlNetModel)
- ):
- if not isinstance(image, list):
- raise TypeError("For multiple controlnets: `image` must be type `list`")
-
- # When `image` is a nested list:
- # (e.g. [[canny_image_1, pose_image_1], [canny_image_2, pose_image_2]])
- elif any(isinstance(i, list) for i in image):
- raise ValueError("A single batch of multiple conditionings are supported at the moment.")
- elif len(image) != len(self.controlnet.nets):
- raise ValueError(
- f"For multiple controlnets: `image` must have the same length as the number of controlnets, but got {len(image)} images and {len(self.controlnet.nets)} ControlNets."
- )
-
- for image_ in image:
- self.check_image(image_, prompt, prompt_embeds)
- else:
- assert False
-
- # Check `controlnet_conditioning_scale`
- if (
- isinstance(self.controlnet, ControlNetModel)
- or is_compiled
- and isinstance(self.controlnet._orig_mod, ControlNetModel)
- ):
- if not isinstance(controlnet_conditioning_scale, float):
- raise TypeError("For single controlnet: `controlnet_conditioning_scale` must be type `float`.")
- elif (
- isinstance(self.controlnet, MultiControlNetModel)
- or is_compiled
- and isinstance(self.controlnet._orig_mod, MultiControlNetModel)
- ):
- if isinstance(controlnet_conditioning_scale, list):
- if any(isinstance(i, list) for i in controlnet_conditioning_scale):
- raise ValueError("A single batch of multiple conditionings are supported at the moment.")
- elif isinstance(controlnet_conditioning_scale, list) and len(controlnet_conditioning_scale) != len(
- self.controlnet.nets
- ):
- raise ValueError(
- "For multiple controlnets: When `controlnet_conditioning_scale` is specified as `list`, it must have"
- " the same length as the number of controlnets"
- )
- else:
- assert False
-
- if len(control_guidance_start) != len(control_guidance_end):
- raise ValueError(
- f"`control_guidance_start` has {len(control_guidance_start)} elements, but `control_guidance_end` has {len(control_guidance_end)} elements. Make sure to provide the same number of elements to each list."
- )
-
- if isinstance(self.controlnet, MultiControlNetModel):
- if len(control_guidance_start) != len(self.controlnet.nets):
- raise ValueError(
- f"`control_guidance_start`: {control_guidance_start} has {len(control_guidance_start)} elements but there are {len(self.controlnet.nets)} controlnets available. Make sure to provide {len(self.controlnet.nets)}."
- )
-
- for start, end in zip(control_guidance_start, control_guidance_end):
- if start >= end:
- raise ValueError(
- f"control guidance start: {start} cannot be larger or equal to control guidance end: {end}."
- )
- if start < 0.0:
- raise ValueError(f"control guidance start: {start} can't be smaller than 0.")
- if end > 1.0:
- raise ValueError(f"control guidance end: {end} can't be larger than 1.0.")
-
- # Copied from diffusers.pipelines.controlnet.pipeline_controlnet.StableDiffusionControlNetPipeline.check_image
- def check_image(self, image, prompt, prompt_embeds):
- image_is_pil = isinstance(image, PIL.Image.Image)
- image_is_tensor = isinstance(image, torch.Tensor)
- image_is_np = isinstance(image, np.ndarray)
- image_is_pil_list = isinstance(image, list) and isinstance(image[0], PIL.Image.Image)
- image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor)
- image_is_np_list = isinstance(image, list) and isinstance(image[0], np.ndarray)
-
- if (
- not image_is_pil
- and not image_is_tensor
- and not image_is_np
- and not image_is_pil_list
- and not image_is_tensor_list
- and not image_is_np_list
- ):
- raise TypeError(
- f"image must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(image)}"
- )
-
- if image_is_pil:
- image_batch_size = 1
- else:
- image_batch_size = len(image)
-
- if prompt is not None and isinstance(prompt, str):
- prompt_batch_size = 1
- elif prompt is not None and isinstance(prompt, list):
- prompt_batch_size = len(prompt)
- elif prompt_embeds is not None:
- prompt_batch_size = prompt_embeds.shape[0]
-
- if image_batch_size != 1 and image_batch_size != prompt_batch_size:
- raise ValueError(
- f"If image batch size is not 1, image batch size must be same as prompt batch size. image batch size: {image_batch_size}, prompt batch size: {prompt_batch_size}"
- )
-
- # Copied from diffusers.pipelines.controlnet.pipeline_controlnet.StableDiffusionControlNetPipeline.prepare_image
- def prepare_control_image(
- self,
- image,
- width,
- height,
- batch_size,
- num_images_per_prompt,
- device,
- dtype,
- do_classifier_free_guidance=False,
- guess_mode=False,
- ):
- image = self.control_image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32)
- image_batch_size = image.shape[0]
-
- if image_batch_size == 1:
- repeat_by = batch_size
- else:
- # image batch size is the same as prompt batch size
- repeat_by = num_images_per_prompt
-
- image = image.repeat_interleave(repeat_by, dim=0)
-
- image = image.to(device=device, dtype=dtype)
-
- if do_classifier_free_guidance and not guess_mode:
- image = torch.cat([image] * 2)
-
- return image
-
- # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint.StableDiffusionInpaintPipeline.prepare_latents
- def prepare_latents(
- self,
- batch_size,
- num_channels_latents,
- height,
- width,
- dtype,
- device,
- generator,
- latents=None,
- image=None,
- timestep=None,
- is_strength_max=True,
- return_noise=False,
- return_image_latents=False,
- ):
- shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
- if isinstance(generator, list) and len(generator) != batch_size:
- raise ValueError(
- f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
- f" size of {batch_size}. Make sure the batch size matches the length of the generators."
- )
-
- if (image is None or timestep is None) and not is_strength_max:
- raise ValueError(
- "Since strength < 1. initial latents are to be initialised as a combination of Image + Noise."
- "However, either the image or the noise timestep has not been provided."
- )
-
- if return_image_latents or (latents is None and not is_strength_max):
- image = image.to(device=device, dtype=dtype)
- image_latents = self._encode_vae_image(image=image, generator=generator)
-
- if latents is None:
- noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
- # if strength is 1. then initialise the latents to noise, else initial to image + noise
- latents = noise if is_strength_max else self.scheduler.add_noise(image_latents, noise, timestep)
- # if pure noise then scale the initial latents by the Scheduler's init sigma
- latents = latents * self.scheduler.init_noise_sigma if is_strength_max else latents
- else:
- noise = latents.to(device)
- latents = noise * self.scheduler.init_noise_sigma
-
- outputs = (latents,)
-
- if return_noise:
- outputs += (noise,)
-
- if return_image_latents:
- outputs += (image_latents,)
-
- return outputs
-
- def _default_height_width(self, height, width, image):
- # NOTE: It is possible that a list of images have different
- # dimensions for each image, so just checking the first image
- # is not _exactly_ correct, but it is simple.
- while isinstance(image, list):
- image = image[0]
-
- if height is None:
- if isinstance(image, PIL.Image.Image):
- height = image.height
- elif isinstance(image, torch.Tensor):
- height = image.shape[2]
-
- height = (height // 8) * 8 # round down to nearest multiple of 8
-
- if width is None:
- if isinstance(image, PIL.Image.Image):
- width = image.width
- elif isinstance(image, torch.Tensor):
- width = image.shape[3]
-
- width = (width // 8) * 8 # round down to nearest multiple of 8
-
- return height, width
-
- # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint.StableDiffusionInpaintPipeline.prepare_mask_latents
- def prepare_mask_latents(
- self, mask, masked_image, batch_size, height, width, dtype, device, generator, do_classifier_free_guidance
- ):
- # resize the mask to latents shape as we concatenate the mask to the latents
- # we do that before converting to dtype to avoid breaking in case we're using cpu_offload
- # and half precision
- mask = torch.nn.functional.interpolate(
- mask, size=(height // self.vae_scale_factor, width // self.vae_scale_factor)
- )
- mask = mask.to(device=device, dtype=dtype)
-
- masked_image = masked_image.to(device=device, dtype=dtype)
- masked_image_latents = self._encode_vae_image(masked_image, generator=generator)
-
- # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method
- if mask.shape[0] < batch_size:
- if not batch_size % mask.shape[0] == 0:
- raise ValueError(
- "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to"
- f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number"
- " of masks that you pass is divisible by the total requested batch size."
- )
- mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1)
- if masked_image_latents.shape[0] < batch_size:
- if not batch_size % masked_image_latents.shape[0] == 0:
- raise ValueError(
- "The passed images and the required batch size don't match. Images are supposed to be duplicated"
- f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed."
- " Make sure the number of images that you pass is divisible by the total requested batch size."
- )
- masked_image_latents = masked_image_latents.repeat(batch_size // masked_image_latents.shape[0], 1, 1, 1)
-
- mask = torch.cat([mask] * 2) if do_classifier_free_guidance else mask
- masked_image_latents = (
- torch.cat([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents
- )
-
- # aligning device to prevent device errors when concating it with the latent model input
- masked_image_latents = masked_image_latents.to(device=device, dtype=dtype)
- return mask, masked_image_latents
-
- # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint.StableDiffusionInpaintPipeline._encode_vae_image
- def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator):
- if isinstance(generator, list):
- image_latents = [
- self.vae.encode(image[i : i + 1]).latent_dist.sample(generator=generator[i])
- for i in range(image.shape[0])
- ]
- image_latents = torch.cat(image_latents, dim=0)
- else:
- image_latents = self.vae.encode(image).latent_dist.sample(generator=generator)
-
- image_latents = self.vae.config.scaling_factor * image_latents
-
- return image_latents
-
- @torch.no_grad()
- def predict_woControl(
- self,
- promptA: Union[str, List[str]] = None,
- promptB: Union[str, List[str]] = None,
- image: Union[torch.FloatTensor, PIL.Image.Image] = None,
- mask_image: Union[torch.FloatTensor, PIL.Image.Image] = None,
- height: Optional[int] = None,
- width: Optional[int] = None,
- strength: float = 1.0,
- tradoff: float = 1.0,
- tradoff_nag: float = 1.0,
- num_inference_steps: int = 50,
- guidance_scale: float = 7.5,
- negative_promptA: Optional[Union[str, List[str]]] = None,
- negative_promptB: Optional[Union[str, List[str]]] = None,
- num_images_per_prompt: Optional[int] = 1,
- eta: float = 0.0,
- generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
- latents: Optional[torch.FloatTensor] = None,
- prompt_embeds: Optional[torch.FloatTensor] = None,
- negative_prompt_embeds: Optional[torch.FloatTensor] = None,
- output_type: Optional[str] = "pil",
- return_dict: bool = True,
- callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
- callback_steps: int = 1,
- cross_attention_kwargs: Optional[Dict[str, Any]] = None,
- task_class: Union[torch.Tensor, float, int] = None,
- ):
- r"""
- The call function to the pipeline for generation.
-
- Args:
- prompt (`str` or `List[str]`, *optional*):
- The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
- image (`PIL.Image.Image`):
- `Image` or tensor representing an image batch to be inpainted (which parts of the image to be masked
- out with `mask_image` and repainted according to `prompt`).
- mask_image (`PIL.Image.Image`):
- `Image` or tensor representing an image batch to mask `image`. White pixels in the mask are repainted
- while black pixels are preserved. If `mask_image` is a PIL image, it is converted to a single channel
- (luminance) before use. If it's a tensor, it should contain one color channel (L) instead of 3, so the
- expected shape would be `(B, H, W, 1)`.
- height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
- The height in pixels of the generated image.
- width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
- The width in pixels of the generated image.
- strength (`float`, *optional*, defaults to 1.0):
- Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a
- starting point and more noise is added the higher the `strength`. The number of denoising steps depends
- on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising
- process runs for the full number of iterations specified in `num_inference_steps`. A value of 1
- essentially ignores `image`.
- num_inference_steps (`int`, *optional*, defaults to 50):
- The number of denoising steps. More denoising steps usually lead to a higher quality image at the
- expense of slower inference. This parameter is modulated by `strength`.
- guidance_scale (`float`, *optional*, defaults to 7.5):
- A higher guidance scale value encourages the model to generate images closely linked to the text
- `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
- negative_prompt (`str` or `List[str]`, *optional*):
- The prompt or prompts to guide what to not include in image generation. If not defined, you need to
- pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
- num_images_per_prompt (`int`, *optional*, defaults to 1):
- The number of images to generate per prompt.
- eta (`float`, *optional*, defaults to 0.0):
- Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
- to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
- generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
- A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
- generation deterministic.
- latents (`torch.FloatTensor`, *optional*):
- Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
- generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
- tensor is generated by sampling using the supplied random `generator`.
- prompt_embeds (`torch.FloatTensor`, *optional*):
- Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
- provided, text embeddings are generated from the `prompt` input argument.
- negative_prompt_embeds (`torch.FloatTensor`, *optional*):
- Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
- not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
- output_type (`str`, *optional*, defaults to `"pil"`):
- The output format of the generated image. Choose between `PIL.Image` or `np.array`.
- return_dict (`bool`, *optional*, defaults to `True`):
- Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
- plain tuple.
- callback (`Callable`, *optional*):
- A function that calls every `callback_steps` steps during inference. The function is called with the
- following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
- callback_steps (`int`, *optional*, defaults to 1):
- The frequency at which the `callback` function is called. If not specified, the callback is called at
- every step.
- cross_attention_kwargs (`dict`, *optional*):
- A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
- [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
-
- Examples:
-
- ```py
- >>> import PIL
- >>> import requests
- >>> import torch
- >>> from io import BytesIO
-
- >>> from diffusers import StableDiffusionInpaintPipeline
-
-
- >>> def download_image(url):
- ... response = requests.get(url)
- ... return PIL.Image.open(BytesIO(response.content)).convert("RGB")
-
-
- >>> img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
- >>> mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"
-
- >>> init_image = download_image(img_url).resize((512, 512))
- >>> mask_image = download_image(mask_url).resize((512, 512))
-
- >>> pipe = StableDiffusionInpaintPipeline.from_pretrained(
- ... "runwayml/stable-diffusion-inpainting", torch_dtype=torch.float16
- ... )
- >>> pipe = pipe.to("cuda")
-
- >>> prompt = "Face of a yellow cat, high resolution, sitting on a park bench"
- >>> image = pipe(prompt=prompt, image=init_image, mask_image=mask_image).images[0]
- ```
-
- Returns:
- [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
- If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
- otherwise a `tuple` is returned where the first element is a list with the generated images and the
- second element is a list of `bool`s indicating whether the corresponding generated image contains
- "not-safe-for-work" (nsfw) content.
- """
- # 0. Default height and width to unet
- height = height or self.unet.config.sample_size * self.vae_scale_factor
- width = width or self.unet.config.sample_size * self.vae_scale_factor
- prompt = promptA
- negative_prompt = negative_promptA
- # 1. Check inputs
- self.check_inputs(
- prompt,
- height,
- width,
- strength,
- callback_steps,
- negative_prompt,
- prompt_embeds,
- negative_prompt_embeds,
- )
-
- # 2. Define call parameters
- if prompt is not None and isinstance(prompt, str):
- batch_size = 1
- elif prompt is not None and isinstance(prompt, list):
- batch_size = len(prompt)
- else:
- batch_size = prompt_embeds.shape[0]
-
- device = self._execution_device
- # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
- # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
- # corresponds to doing no classifier free guidance.
- do_classifier_free_guidance = guidance_scale > 1.0
-
- # 3. Encode input prompt
- text_encoder_lora_scale = (
- cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
- )
- prompt_embeds = self._encode_prompt(
- promptA,
- promptB,
- tradoff,
- device,
- num_images_per_prompt,
- do_classifier_free_guidance,
- negative_promptA,
- negative_promptB,
- tradoff_nag,
- prompt_embeds=prompt_embeds,
- negative_prompt_embeds=negative_prompt_embeds,
- lora_scale=text_encoder_lora_scale,
- )
-
- # 4. set timesteps
- self.scheduler.set_timesteps(num_inference_steps, device=device)
- timesteps, num_inference_steps = self.get_timesteps(
- num_inference_steps=num_inference_steps, strength=strength, device=device
- )
- # check that number of inference steps is not < 1 - as this doesn't make sense
- if num_inference_steps < 1:
- raise ValueError(
- f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline"
- f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline."
- )
- # at which timestep to set the initial noise (n.b. 50% if strength is 0.5)
- latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
- # create a boolean to check if the strength is set to 1. if so then initialise the latents with pure noise
- is_strength_max = strength == 1.0
-
- # 5. Preprocess mask and image
- mask, masked_image, init_image = prepare_mask_and_masked_image(
- image, mask_image, height, width, return_image=True
- )
- mask_condition = mask.clone()
-
- # 6. Prepare latent variables
- num_channels_latents = self.vae.config.latent_channels
- num_channels_unet = self.unet.config.in_channels
- return_image_latents = num_channels_unet == 4
-
- latents_outputs = self.prepare_latents(
- batch_size * num_images_per_prompt,
- num_channels_latents,
- height,
- width,
- prompt_embeds.dtype,
- device,
- generator,
- latents,
- image=init_image,
- timestep=latent_timestep,
- is_strength_max=is_strength_max,
- return_noise=True,
- return_image_latents=return_image_latents,
- )
-
- if return_image_latents:
- latents, noise, image_latents = latents_outputs
- else:
- latents, noise = latents_outputs
-
- # 7. Prepare mask latent variables
- mask, masked_image_latents = self.prepare_mask_latents(
- mask,
- masked_image,
- batch_size * num_images_per_prompt,
- height,
- width,
- prompt_embeds.dtype,
- device,
- generator,
- do_classifier_free_guidance,
- )
-
- # 8. Check that sizes of mask, masked image and latents match
- if num_channels_unet == 9:
- # default case for runwayml/stable-diffusion-inpainting
- num_channels_mask = mask.shape[1]
- num_channels_masked_image = masked_image_latents.shape[1]
- if num_channels_latents + num_channels_mask + num_channels_masked_image != self.unet.config.in_channels:
- raise ValueError(
- f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
- f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
- f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
- f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
- " `pipeline.unet` or your `mask_image` or `image` input."
- )
- elif num_channels_unet != 4:
- raise ValueError(
- f"The unet {self.unet.__class__} should have either 4 or 9 input channels, not {self.unet.config.in_channels}."
- )
-
- # 9. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
- extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
-
- # 10. Denoising loop
- num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
- with self.progress_bar(total=num_inference_steps) as progress_bar:
- for i, t in enumerate(timesteps):
- # expand the latents if we are doing classifier free guidance
- latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
-
- # concat latents, mask, masked_image_latents in the channel dimension
- latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
-
- if num_channels_unet == 9:
- latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1)
-
- # predict the noise residual
- if task_class is not None:
- noise_pred = self.unet(
- sample = latent_model_input,
- timestep = t,
- encoder_hidden_states=prompt_embeds,
- cross_attention_kwargs=cross_attention_kwargs,
- return_dict=False,
- task_class = task_class,
- )[0]
- else:
- noise_pred = self.unet(
- latent_model_input,
- t,
- encoder_hidden_states=prompt_embeds,
- cross_attention_kwargs=cross_attention_kwargs,
- return_dict=False,
- )[0]
-
- # perform guidance
- if do_classifier_free_guidance:
- noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
- noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
-
- # compute the previous noisy sample x_t -> x_t-1
- latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
-
- if num_channels_unet == 4:
- init_latents_proper = image_latents[:1]
- init_mask = mask[:1]
-
- if i < len(timesteps) - 1:
- noise_timestep = timesteps[i + 1]
- init_latents_proper = self.scheduler.add_noise(
- init_latents_proper, noise, torch.tensor([noise_timestep])
- )
-
- latents = (1 - init_mask) * init_latents_proper + init_mask * latents
-
- # call the callback, if provided
- if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
- progress_bar.update()
- if callback is not None and i % callback_steps == 0:
- callback(i, t, latents)
-
- if not output_type == "latent":
- condition_kwargs = {}
- if isinstance(self.vae, AsymmetricAutoencoderKL):
- init_image = init_image.to(device=device, dtype=masked_image_latents.dtype)
- init_image_condition = init_image.clone()
- init_image = self._encode_vae_image(init_image, generator=generator)
- mask_condition = mask_condition.to(device=device, dtype=masked_image_latents.dtype)
- condition_kwargs = {"image": init_image_condition, "mask": mask_condition}
- image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, **condition_kwargs)[0]
- image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
- else:
- image = latents
- has_nsfw_concept = None
-
- if has_nsfw_concept is None:
- do_denormalize = [True] * image.shape[0]
- else:
- do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
-
- image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
-
- # Offload last model to CPU
- if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
- self.final_offload_hook.offload()
-
- if not return_dict:
- return (image, has_nsfw_concept)
-
- return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
-
-
- @torch.no_grad()
- @replace_example_docstring(EXAMPLE_DOC_STRING)
- def __call__(
- self,
- promptA: Union[str, List[str]] = None,
- promptB: Union[str, List[str]] = None,
- image: Union[torch.Tensor, PIL.Image.Image] = None,
- mask_image: Union[torch.Tensor, PIL.Image.Image] = None,
- control_image: Union[
- torch.FloatTensor,
- PIL.Image.Image,
- np.ndarray,
- List[torch.FloatTensor],
- List[PIL.Image.Image],
- List[np.ndarray],
- ] = None,
- height: Optional[int] = None,
- width: Optional[int] = None,
- strength: float = 1.0,
- tradoff: float = 1.0,
- tradoff_nag: float = 1.0,
- num_inference_steps: int = 50,
- guidance_scale: float = 7.5,
- negative_promptA: Optional[Union[str, List[str]]] = None,
- negative_promptB: Optional[Union[str, List[str]]] = None,
- num_images_per_prompt: Optional[int] = 1,
- eta: float = 0.0,
- generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
- latents: Optional[torch.FloatTensor] = None,
- prompt_embeds: Optional[torch.FloatTensor] = None,
- negative_prompt_embeds: Optional[torch.FloatTensor] = None,
- output_type: Optional[str] = "pil",
- return_dict: bool = True,
- callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
- callback_steps: int = 1,
- cross_attention_kwargs: Optional[Dict[str, Any]] = None,
- controlnet_conditioning_scale: Union[float, List[float]] = 0.5,
- guess_mode: bool = False,
- control_guidance_start: Union[float, List[float]] = 0.0,
- control_guidance_end: Union[float, List[float]] = 1.0,
- ):
- r"""
- Function invoked when calling the pipeline for generation.
-
- Args:
- prompt (`str` or `List[str]`, *optional*):
- The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
- instead.
- image (`torch.FloatTensor`, `PIL.Image.Image`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`,
- `List[List[torch.FloatTensor]]`, or `List[List[PIL.Image.Image]]`):
- The ControlNet input condition. ControlNet uses this input condition to generate guidance to Unet. If
- the type is specified as `Torch.FloatTensor`, it is passed to ControlNet as is. `PIL.Image.Image` can
- also be accepted as an image. The dimensions of the output image defaults to `image`'s dimensions. If
- height and/or width are passed, `image` is resized according to them. If multiple ControlNets are
- specified in init, images must be passed as a list such that each element of the list can be correctly
- batched for input to a single controlnet.
- height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
- The height in pixels of the generated image.
- width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
- The width in pixels of the generated image.
- strength (`float`, *optional*, defaults to 1.):
- Conceptually, indicates how much to transform the masked portion of the reference `image`. Must be
- between 0 and 1. `image` will be used as a starting point, adding more noise to it the larger the
- `strength`. The number of denoising steps depends on the amount of noise initially added. When
- `strength` is 1, added noise will be maximum and the denoising process will run for the full number of
- iterations specified in `num_inference_steps`. A value of 1, therefore, essentially ignores the masked
- portion of the reference `image`.
- num_inference_steps (`int`, *optional*, defaults to 50):
- The number of denoising steps. More denoising steps usually lead to a higher quality image at the
- expense of slower inference.
- guidance_scale (`float`, *optional*, defaults to 7.5):
- Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
- `guidance_scale` is defined as `w` of equation 2. of [Imagen
- Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
- 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
- usually at the expense of lower image quality.
- negative_prompt (`str` or `List[str]`, *optional*):
- The prompt or prompts not to guide the image generation. If not defined, one has to pass
- `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
- less than `1`).
- num_images_per_prompt (`int`, *optional*, defaults to 1):
- The number of images to generate per prompt.
- eta (`float`, *optional*, defaults to 0.0):
- Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
- [`schedulers.DDIMScheduler`], will be ignored for others.
- generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
- One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
- to make generation deterministic.
- latents (`torch.FloatTensor`, *optional*):
- Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
- generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
- tensor will ge generated by sampling using the supplied random `generator`.
- prompt_embeds (`torch.FloatTensor`, *optional*):
- Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
- provided, text embeddings will be generated from `prompt` input argument.
- negative_prompt_embeds (`torch.FloatTensor`, *optional*):
- Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
- weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
- argument.
- output_type (`str`, *optional*, defaults to `"pil"`):
- The output format of the generate image. Choose between
- [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
- return_dict (`bool`, *optional*, defaults to `True`):
- Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
- plain tuple.
- callback (`Callable`, *optional*):
- A function that will be called every `callback_steps` steps during inference. The function will be
- called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
- callback_steps (`int`, *optional*, defaults to 1):
- The frequency at which the `callback` function will be called. If not specified, the callback will be
- called at every step.
- cross_attention_kwargs (`dict`, *optional*):
- A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
- `self.processor` in
- [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
- controlnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 0.5):
- The outputs of the controlnet are multiplied by `controlnet_conditioning_scale` before they are added
- to the residual in the original unet. If multiple ControlNets are specified in init, you can set the
- corresponding scale as a list. Note that by default, we use a smaller conditioning scale for inpainting
- than for [`~StableDiffusionControlNetPipeline.__call__`].
- guess_mode (`bool`, *optional*, defaults to `False`):
- In this mode, the ControlNet encoder will try best to recognize the content of the input image even if
- you remove all prompts. The `guidance_scale` between 3.0 and 5.0 is recommended.
- control_guidance_start (`float` or `List[float]`, *optional*, defaults to 0.0):
- The percentage of total steps at which the controlnet starts applying.
- control_guidance_end (`float` or `List[float]`, *optional*, defaults to 1.0):
- The percentage of total steps at which the controlnet stops applying.
-
- Examples:
-
- Returns:
- [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
- [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
- When returning a tuple, the first element is a list with the generated images, and the second element is a
- list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
- (nsfw) content, according to the `safety_checker`.
- """
- controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet
-
- # 0. Default height and width to unet
- height, width = self._default_height_width(height, width, image)
-
- prompt = promptA
- negative_prompt = negative_promptA
-
- # align format for control guidance
- if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list):
- control_guidance_start = len(control_guidance_end) * [control_guidance_start]
- elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list):
- control_guidance_end = len(control_guidance_start) * [control_guidance_end]
- elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list):
- mult = len(controlnet.nets) if isinstance(controlnet, MultiControlNetModel) else 1
- control_guidance_start, control_guidance_end = mult * [control_guidance_start], mult * [
- control_guidance_end
- ]
-
- # 1. Check inputs. Raise error if not correct
- self.check_inputs(
- prompt,
- control_image,
- height,
- width,
- callback_steps,
- negative_prompt,
- prompt_embeds,
- negative_prompt_embeds,
- controlnet_conditioning_scale,
- control_guidance_start,
- control_guidance_end,
- )
-
- # 2. Define call parameters
- if prompt is not None and isinstance(prompt, str):
- batch_size = 1
- elif prompt is not None and isinstance(prompt, list):
- batch_size = len(prompt)
- else:
- batch_size = prompt_embeds.shape[0]
-
- device = self._execution_device
- # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
- # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
- # corresponds to doing no classifier free guidance.
- do_classifier_free_guidance = guidance_scale > 1.0
-
- if isinstance(controlnet, MultiControlNetModel) and isinstance(controlnet_conditioning_scale, float):
- controlnet_conditioning_scale = [controlnet_conditioning_scale] * len(controlnet.nets)
-
- global_pool_conditions = (
- controlnet.config.global_pool_conditions
- if isinstance(controlnet, ControlNetModel)
- else controlnet.nets[0].config.global_pool_conditions
- )
- guess_mode = guess_mode or global_pool_conditions
-
- # 3. Encode input prompt
- text_encoder_lora_scale = (
- cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
- )
- prompt_embeds = self._encode_prompt(
- promptA,
- promptB,
- tradoff,
- device,
- num_images_per_prompt,
- do_classifier_free_guidance,
- negative_promptA,
- negative_promptB,
- tradoff_nag,
- prompt_embeds=prompt_embeds,
- negative_prompt_embeds=negative_prompt_embeds,
- lora_scale=text_encoder_lora_scale,
- )
-
- # 4. Prepare image
- if isinstance(controlnet, ControlNetModel):
- control_image = self.prepare_control_image(
- image=control_image,
- width=width,
- height=height,
- batch_size=batch_size * num_images_per_prompt,
- num_images_per_prompt=num_images_per_prompt,
- device=device,
- dtype=controlnet.dtype,
- do_classifier_free_guidance=do_classifier_free_guidance,
- guess_mode=guess_mode,
- )
- elif isinstance(controlnet, MultiControlNetModel):
- control_images = []
-
- for control_image_ in control_image:
- control_image_ = self.prepare_control_image(
- image=control_image_,
- width=width,
- height=height,
- batch_size=batch_size * num_images_per_prompt,
- num_images_per_prompt=num_images_per_prompt,
- device=device,
- dtype=controlnet.dtype,
- do_classifier_free_guidance=do_classifier_free_guidance,
- guess_mode=guess_mode,
- )
-
- control_images.append(control_image_)
-
- control_image = control_images
- else:
- assert False
-
- # 4. Preprocess mask and image - resizes image and mask w.r.t height and width
- mask, masked_image, init_image = prepare_mask_and_masked_image(
- image, mask_image, height, width, return_image=True
- )
-
- # 5. Prepare timesteps
- self.scheduler.set_timesteps(num_inference_steps, device=device)
- timesteps, num_inference_steps = self.get_timesteps(
- num_inference_steps=num_inference_steps, strength=strength, device=device
- )
- # at which timestep to set the initial noise (n.b. 50% if strength is 0.5)
- latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
- # create a boolean to check if the strength is set to 1. if so then initialise the latents with pure noise
- is_strength_max = strength == 1.0
-
- # 6. Prepare latent variables
- num_channels_latents = self.vae.config.latent_channels
- num_channels_unet = self.unet.config.in_channels
- return_image_latents = num_channels_unet == 4
- latents_outputs = self.prepare_latents(
- batch_size * num_images_per_prompt,
- num_channels_latents,
- height,
- width,
- prompt_embeds.dtype,
- device,
- generator,
- latents,
- image=init_image,
- timestep=latent_timestep,
- is_strength_max=is_strength_max,
- return_noise=True,
- return_image_latents=return_image_latents,
- )
-
- if return_image_latents:
- latents, noise, image_latents = latents_outputs
- else:
- latents, noise = latents_outputs
-
- # 7. Prepare mask latent variables
- mask, masked_image_latents = self.prepare_mask_latents(
- mask,
- masked_image,
- batch_size * num_images_per_prompt,
- height,
- width,
- prompt_embeds.dtype,
- device,
- generator,
- do_classifier_free_guidance,
- )
-
- # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
- extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
-
- # 7.1 Create tensor stating which controlnets to keep
- controlnet_keep = []
- for i in range(len(timesteps)):
- keeps = [
- 1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e)
- for s, e in zip(control_guidance_start, control_guidance_end)
- ]
- controlnet_keep.append(keeps[0] if isinstance(controlnet, ControlNetModel) else keeps)
-
- # 8. Denoising loop
- num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
- with self.progress_bar(total=num_inference_steps) as progress_bar:
- for i, t in enumerate(timesteps):
- # expand the latents if we are doing classifier free guidance
- latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
- latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
-
- # controlnet(s) inference
- if guess_mode and do_classifier_free_guidance:
- # Infer ControlNet only for the conditional batch.
- control_model_input = latents
- control_model_input = self.scheduler.scale_model_input(control_model_input, t)
- controlnet_prompt_embeds = prompt_embeds.chunk(2)[1]
- else:
- control_model_input = latent_model_input
- controlnet_prompt_embeds = prompt_embeds
-
- if isinstance(controlnet_keep[i], list):
- cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])]
- else:
- controlnet_cond_scale = controlnet_conditioning_scale
- if isinstance(controlnet_cond_scale, list):
- controlnet_cond_scale = controlnet_cond_scale[0]
- cond_scale = controlnet_cond_scale * controlnet_keep[i]
-
- down_block_res_samples, mid_block_res_sample = self.controlnet(
- control_model_input,
- t,
- encoder_hidden_states=controlnet_prompt_embeds,
- controlnet_cond=control_image,
- conditioning_scale=cond_scale,
- guess_mode=guess_mode,
- return_dict=False,
- )
-
- if guess_mode and do_classifier_free_guidance:
- # Infered ControlNet only for the conditional batch.
- # To apply the output of ControlNet to both the unconditional and conditional batches,
- # add 0 to the unconditional batch to keep it unchanged.
- down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples]
- mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample])
-
- # predict the noise residual
- if num_channels_unet == 9:
- latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1)
-
- noise_pred = self.unet(
- latent_model_input,
- t,
- encoder_hidden_states=prompt_embeds,
- cross_attention_kwargs=cross_attention_kwargs,
- down_block_additional_residuals=down_block_res_samples,
- mid_block_additional_residual=mid_block_res_sample,
- return_dict=False,
- )[0]
-
- # perform guidance
- if do_classifier_free_guidance:
- noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
- noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
-
- # compute the previous noisy sample x_t -> x_t-1
- latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
-
- if num_channels_unet == 4:
- init_latents_proper = image_latents[:1]
- init_mask = mask[:1]
-
- if i < len(timesteps) - 1:
- noise_timestep = timesteps[i + 1]
- init_latents_proper = self.scheduler.add_noise(
- init_latents_proper, noise, torch.tensor([noise_timestep])
- )
-
- latents = (1 - init_mask) * init_latents_proper + init_mask * latents
-
- # call the callback, if provided
- if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
- progress_bar.update()
- if callback is not None and i % callback_steps == 0:
- callback(i, t, latents)
-
- # If we do sequential model offloading, let's offload unet and controlnet
- # manually for max memory savings
- if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
- self.unet.to("cpu")
- self.controlnet.to("cpu")
- torch.cuda.empty_cache()
-
- if not output_type == "latent":
- image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
- image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
- else:
- image = latents
- has_nsfw_concept = None
-
- if has_nsfw_concept is None:
- do_denormalize = [True] * image.shape[0]
- else:
- do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
-
- image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
-
- # Offload last model to CPU
- if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
- self.final_offload_hook.offload()
-
- if not return_dict:
- return (image, has_nsfw_concept)
-
- return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/iopaint/model/power_paint/power_paint_v2.py b/iopaint/model/power_paint/power_paint_v2.py
new file mode 100644
index 0000000..55ea5e1
--- /dev/null
+++ b/iopaint/model/power_paint/power_paint_v2.py
@@ -0,0 +1,141 @@
+import PIL.Image
+import cv2
+import torch
+from loguru import logger
+from transformers import CLIPTextModel, CLIPTokenizer
+import numpy as np
+
+from ..base import DiffusionInpaintModel
+from ..helper.cpu_text_encoder import CPUTextEncoderWrapper
+from ..utils import (
+ get_torch_dtype,
+ enable_low_mem,
+ is_local_files_only,
+ handle_from_pretrained_exceptions,
+)
+from .powerpaint_tokenizer import task_to_prompt
+from iopaint.schema import InpaintRequest
+from .v2.BrushNet_CA import BrushNetModel
+from .v2.unet_2d_condition import UNet2DConditionModel
+
+
+class PowerPaintV2(DiffusionInpaintModel):
+ pad_mod = 8
+ min_size = 512
+ lcm_lora_id = "latent-consistency/lcm-lora-sdv1-5"
+ hf_model_id = "Sanster/PowerPaint_v2"
+
+ def init_model(self, device: torch.device, **kwargs):
+ from .v2.pipeline_PowerPaint_Brushnet_CA import (
+ StableDiffusionPowerPaintBrushNetPipeline,
+ )
+ from .powerpaint_tokenizer import PowerPaintTokenizer
+
+ use_gpu, torch_dtype = get_torch_dtype(device, kwargs.get("no_half", False))
+ model_kwargs = {"local_files_only": is_local_files_only(**kwargs)}
+ if kwargs["disable_nsfw"] or kwargs.get("cpu_offload", False):
+ logger.info("Disable Stable Diffusion Model NSFW checker")
+ model_kwargs.update(
+ dict(
+ safety_checker=None,
+ feature_extractor=None,
+ requires_safety_checker=False,
+ )
+ )
+
+ text_encoder_brushnet = CLIPTextModel.from_pretrained(
+ self.hf_model_id,
+ subfolder="text_encoder_brushnet",
+ variant="fp16",
+ torch_dtype=torch_dtype,
+ local_files_only=model_kwargs["local_files_only"],
+ )
+ unet = handle_from_pretrained_exceptions(
+ UNet2DConditionModel.from_pretrained,
+ pretrained_model_name_or_path=self.model_id_or_path,
+ subfolder="unet",
+ variant="fp16",
+ torch_dtype=torch_dtype,
+ local_files_only=model_kwargs["local_files_only"],
+ )
+ brushnet = BrushNetModel.from_pretrained(
+ self.hf_model_id,
+ subfolder="PowerPaint_Brushnet",
+ variant="fp16",
+ torch_dtype=torch_dtype,
+ local_files_only=model_kwargs["local_files_only"],
+ )
+ pipe = handle_from_pretrained_exceptions(
+ StableDiffusionPowerPaintBrushNetPipeline.from_pretrained,
+ pretrained_model_name_or_path=self.model_id_or_path,
+ torch_dtype=torch_dtype,
+ unet=unet,
+ brushnet=brushnet,
+ text_encoder_brushnet=text_encoder_brushnet,
+ variant="fp16",
+ **model_kwargs,
+ )
+ pipe.tokenizer = PowerPaintTokenizer(
+ CLIPTokenizer.from_pretrained(self.hf_model_id, subfolder="tokenizer")
+ )
+ self.model = pipe
+
+ enable_low_mem(self.model, kwargs.get("low_mem", False))
+
+ if kwargs.get("cpu_offload", False) and use_gpu:
+ logger.info("Enable sequential cpu offload")
+ self.model.enable_sequential_cpu_offload(gpu_id=0)
+ else:
+ self.model = self.model.to(device)
+ if kwargs["sd_cpu_textencoder"]:
+ logger.info("Run Stable Diffusion TextEncoder on CPU")
+ self.model.text_encoder = CPUTextEncoderWrapper(
+ self.model.text_encoder, torch_dtype
+ )
+
+ self.callback = kwargs.pop("callback", None)
+
+ def forward(self, image, mask, config: InpaintRequest):
+ """Input image and output image have same size
+ image: [H, W, C] RGB
+ mask: [H, W, 1] 255 means area to repaint
+ return: BGR IMAGE
+ """
+ self.set_scheduler(config)
+
+ image = image * (1 - mask / 255.0)
+ img_h, img_w = image.shape[:2]
+
+ image = PIL.Image.fromarray(image.astype(np.uint8))
+ mask = PIL.Image.fromarray(mask[:, :, -1], mode="L").convert("RGB")
+
+ promptA, promptB, negative_promptA, negative_promptB = task_to_prompt(
+ config.powerpaint_task
+ )
+
+ output = self.model(
+ image=image,
+ mask=mask,
+ promptA=promptA,
+ promptB=promptB,
+ promptU=config.prompt,
+ tradoff=config.fitting_degree,
+ tradoff_nag=config.fitting_degree,
+ negative_promptA=negative_promptA,
+ negative_promptB=negative_promptB,
+ negative_promptU=config.negative_prompt,
+ num_inference_steps=config.sd_steps,
+ # strength=config.sd_strength,
+ brushnet_conditioning_scale=1.0,
+ guidance_scale=config.sd_guidance_scale,
+ output_type="np",
+ callback=self.callback,
+ height=img_h,
+ width=img_w,
+ generator=torch.manual_seed(config.sd_seed),
+ callback_steps=1,
+ ).images[0]
+
+ output = (output * 255).round().astype("uint8")
+ output = cv2.cvtColor(output, cv2.COLOR_RGB2BGR)
+ return output
diff --git a/iopaint/model/power_paint/powerpaint_tokenizer.py b/iopaint/model/power_paint/powerpaint_tokenizer.py
index 39d5cb7..53a68c9 100644
--- a/iopaint/model/power_paint/powerpaint_tokenizer.py
+++ b/iopaint/model/power_paint/powerpaint_tokenizer.py
@@ -1,8 +1,6 @@
-import torch
-import torch.nn as nn
import copy
import random
-from typing import Any, List, Optional, Union
+from typing import Any, List, Union
from transformers import CLIPTokenizer
from iopaint.schema import PowerPaintTask
@@ -14,6 +12,11 @@ def add_task_to_prompt(prompt, negative_prompt, task: PowerPaintTask):
promptB = prompt + " P_ctxt"
negative_promptA = negative_prompt + " P_obj"
negative_promptB = negative_prompt + " P_obj"
+ elif task == PowerPaintTask.context_aware:
+ promptA = prompt + " P_ctxt"
+ promptB = prompt + " P_ctxt"
+ negative_promptA = negative_prompt
+ negative_promptB = negative_prompt
elif task == PowerPaintTask.shape_guided:
promptA = prompt + " P_shape"
promptB = prompt + " P_ctxt"
@@ -33,6 +36,18 @@ def add_task_to_prompt(prompt, negative_prompt, task: PowerPaintTask):
return promptA, promptB, negative_promptA, negative_promptB
+def task_to_prompt(task: PowerPaintTask):
+ promptA, promptB, negative_promptA, negative_promptB = add_task_to_prompt(
+ "", "", task
+ )
+ return (
+ promptA.strip(),
+ promptB.strip(),
+ negative_promptA.strip(),
+ negative_promptB.strip(),
+ )
+
+
class PowerPaintTokenizer:
def __init__(self, tokenizer: CLIPTokenizer):
self.wrapped = tokenizer
@@ -237,304 +252,3 @@ class PowerPaintTokenizer:
return text
replaced_text = self.replace_text_with_placeholder_tokens(text)
return replaced_text
-
-
-class EmbeddingLayerWithFixes(nn.Module):
- """The revised embedding layer to support external embeddings. This design
- of this class is inspired by https://github.com/AUTOMATIC1111/stable-
- diffusion-webui/blob/22bcc7be428c94e9408f589966c2040187245d81/modules/sd_hi
- jack.py#L224 # noqa.
-
- Args:
- wrapped (nn.Emebdding): The embedding layer to be wrapped.
- external_embeddings (Union[dict, List[dict]], optional): The external
- embeddings added to this layer. Defaults to None.
- """
-
- def __init__(
- self,
- wrapped: nn.Embedding,
- external_embeddings: Optional[Union[dict, List[dict]]] = None,
- ):
- super().__init__()
- self.wrapped = wrapped
- self.num_embeddings = wrapped.weight.shape[0]
-
- self.external_embeddings = []
- if external_embeddings:
- self.add_embeddings(external_embeddings)
-
- self.trainable_embeddings = nn.ParameterDict()
-
- @property
- def weight(self):
- """Get the weight of wrapped embedding layer."""
- return self.wrapped.weight
-
- def check_duplicate_names(self, embeddings: List[dict]):
- """Check whether duplicate names exist in list of 'external
- embeddings'.
-
- Args:
- embeddings (List[dict]): A list of embedding to be check.
- """
- names = [emb["name"] for emb in embeddings]
- assert len(names) == len(set(names)), (
- "Found duplicated names in 'external_embeddings'. Name list: " f"'{names}'"
- )
-
- def check_ids_overlap(self, embeddings):
- """Check whether overlap exist in token ids of 'external_embeddings'.
-
- Args:
- embeddings (List[dict]): A list of embedding to be check.
- """
- ids_range = [[emb["start"], emb["end"], emb["name"]] for emb in embeddings]
- ids_range.sort() # sort by 'start'
- # check if 'end' has overlapping
- for idx in range(len(ids_range) - 1):
- name1, name2 = ids_range[idx][-1], ids_range[idx + 1][-1]
- assert ids_range[idx][1] <= ids_range[idx + 1][0], (
- f"Found ids overlapping between embeddings '{name1}' " f"and '{name2}'."
- )
-
- def add_embeddings(self, embeddings: Optional[Union[dict, List[dict]]]):
- """Add external embeddings to this layer.
-
- Use case:
-
- >>> 1. Add token to tokenizer and get the token id.
- >>> tokenizer = TokenizerWrapper('openai/clip-vit-base-patch32')
- >>> # 'how much' in kiswahili
- >>> tokenizer.add_placeholder_tokens('ngapi', num_vec_per_token=4)
- >>>
- >>> 2. Add external embeddings to the model.
- >>> new_embedding = {
- >>> 'name': 'ngapi', # 'how much' in kiswahili
- >>> 'embedding': torch.ones(1, 15) * 4,
- >>> 'start': tokenizer.get_token_info('kwaheri')['start'],
- >>> 'end': tokenizer.get_token_info('kwaheri')['end'],
- >>> 'trainable': False # if True, will registry as a parameter
- >>> }
- >>> embedding_layer = nn.Embedding(10, 15)
- >>> embedding_layer_wrapper = EmbeddingLayerWithFixes(embedding_layer)
- >>> embedding_layer_wrapper.add_embeddings(new_embedding)
- >>>
- >>> 3. Forward tokenizer and embedding layer!
- >>> input_text = ['hello, ngapi!', 'hello my friend, ngapi?']
- >>> input_ids = tokenizer(
- >>> input_text, padding='max_length', truncation=True,
- >>> return_tensors='pt')['input_ids']
- >>> out_feat = embedding_layer_wrapper(input_ids)
- >>>
- >>> 4. Let's validate the result!
- >>> assert (out_feat[0, 3: 7] == 2.3).all()
- >>> assert (out_feat[2, 5: 9] == 2.3).all()
-
- Args:
- embeddings (Union[dict, list[dict]]): The external embeddings to
- be added. Each dict must contain the following 4 fields: 'name'
- (the name of this embedding), 'embedding' (the embedding
- tensor), 'start' (the start token id of this embedding), 'end'
- (the end token id of this embedding). For example:
- `{name: NAME, start: START, end: END, embedding: torch.Tensor}`
- """
- if isinstance(embeddings, dict):
- embeddings = [embeddings]
-
- self.external_embeddings += embeddings
- self.check_duplicate_names(self.external_embeddings)
- self.check_ids_overlap(self.external_embeddings)
-
- # set for trainable
- added_trainable_emb_info = []
- for embedding in embeddings:
- trainable = embedding.get("trainable", False)
- if trainable:
- name = embedding["name"]
- embedding["embedding"] = torch.nn.Parameter(embedding["embedding"])
- self.trainable_embeddings[name] = embedding["embedding"]
- added_trainable_emb_info.append(name)
-
- added_emb_info = [emb["name"] for emb in embeddings]
- added_emb_info = ", ".join(added_emb_info)
- print(f"Successfully add external embeddings: {added_emb_info}.", "current")
-
- if added_trainable_emb_info:
- added_trainable_emb_info = ", ".join(added_trainable_emb_info)
- print(
- "Successfully add trainable external embeddings: "
- f"{added_trainable_emb_info}",
- "current",
- )
-
- def replace_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
- """Replace external input ids to 0.
-
- Args:
- input_ids (torch.Tensor): The input ids to be replaced.
-
- Returns:
- torch.Tensor: The replaced input ids.
- """
- input_ids_fwd = input_ids.clone()
- input_ids_fwd[input_ids_fwd >= self.num_embeddings] = 0
- return input_ids_fwd
-
- def replace_embeddings(
- self, input_ids: torch.Tensor, embedding: torch.Tensor, external_embedding: dict
- ) -> torch.Tensor:
- """Replace external embedding to the embedding layer. Noted that, in
- this function we use `torch.cat` to avoid inplace modification.
-
- Args:
- input_ids (torch.Tensor): The original token ids. Shape like
- [LENGTH, ].
- embedding (torch.Tensor): The embedding of token ids after
- `replace_input_ids` function.
- external_embedding (dict): The external embedding to be replaced.
-
- Returns:
- torch.Tensor: The replaced embedding.
- """
- new_embedding = []
-
- name = external_embedding["name"]
- start = external_embedding["start"]
- end = external_embedding["end"]
- target_ids_to_replace = [i for i in range(start, end)]
- ext_emb = external_embedding["embedding"]
-
- # do not need to replace
- if not (input_ids == start).any():
- return embedding
-
- # start replace
- s_idx, e_idx = 0, 0
- while e_idx < len(input_ids):
- if input_ids[e_idx] == start:
- if e_idx != 0:
- # add embedding do not need to replace
- new_embedding.append(embedding[s_idx:e_idx])
-
- # check if the next embedding need to replace is valid
- actually_ids_to_replace = [
- int(i) for i in input_ids[e_idx : e_idx + end - start]
- ]
- assert actually_ids_to_replace == target_ids_to_replace, (
- f"Invalid 'input_ids' in position: {s_idx} to {e_idx}. "
- f"Expect '{target_ids_to_replace}' for embedding "
- f"'{name}' but found '{actually_ids_to_replace}'."
- )
-
- new_embedding.append(ext_emb)
-
- s_idx = e_idx + end - start
- e_idx = s_idx + 1
- else:
- e_idx += 1
-
- if e_idx == len(input_ids):
- new_embedding.append(embedding[s_idx:e_idx])
-
- return torch.cat(new_embedding, dim=0)
-
- def forward(
- self, input_ids: torch.Tensor, external_embeddings: Optional[List[dict]] = None
- ):
- """The forward function.
-
- Args:
- input_ids (torch.Tensor): The token ids shape like [bz, LENGTH] or
- [LENGTH, ].
- external_embeddings (Optional[List[dict]]): The external
- embeddings. If not passed, only `self.external_embeddings`
- will be used. Defaults to None.
-
- input_ids: shape like [bz, LENGTH] or [LENGTH].
- """
- assert input_ids.ndim in [1, 2]
- if input_ids.ndim == 1:
- input_ids = input_ids.unsqueeze(0)
-
- if external_embeddings is None and not self.external_embeddings:
- return self.wrapped(input_ids)
-
- input_ids_fwd = self.replace_input_ids(input_ids)
- inputs_embeds = self.wrapped(input_ids_fwd)
-
- vecs = []
-
- if external_embeddings is None:
- external_embeddings = []
- elif isinstance(external_embeddings, dict):
- external_embeddings = [external_embeddings]
- embeddings = self.external_embeddings + external_embeddings
-
- for input_id, embedding in zip(input_ids, inputs_embeds):
- new_embedding = embedding
- for external_embedding in embeddings:
- new_embedding = self.replace_embeddings(
- input_id, new_embedding, external_embedding
- )
- vecs.append(new_embedding)
-
- return torch.stack(vecs)
-
-
-def add_tokens(
- tokenizer,
- text_encoder,
- placeholder_tokens: list,
- initialize_tokens: list = None,
- num_vectors_per_token: int = 1,
-):
- """Add token for training.
-
- # TODO: support add tokens as dict, then we can load pretrained tokens.
- """
- if initialize_tokens is not None:
- assert len(initialize_tokens) == len(
- placeholder_tokens
- ), "placeholder_token should be the same length as initialize_token"
- for ii in range(len(placeholder_tokens)):
- tokenizer.add_placeholder_token(
- placeholder_tokens[ii], num_vec_per_token=num_vectors_per_token
- )
-
- # text_encoder.set_embedding_layer()
- embedding_layer = text_encoder.text_model.embeddings.token_embedding
- text_encoder.text_model.embeddings.token_embedding = EmbeddingLayerWithFixes(
- embedding_layer
- )
- embedding_layer = text_encoder.text_model.embeddings.token_embedding
-
- assert embedding_layer is not None, (
- "Do not support get embedding layer for current text encoder. "
- "Please check your configuration."
- )
- initialize_embedding = []
- if initialize_tokens is not None:
- for ii in range(len(placeholder_tokens)):
- init_id = tokenizer(initialize_tokens[ii]).input_ids[1]
- temp_embedding = embedding_layer.weight[init_id]
- initialize_embedding.append(
- temp_embedding[None, ...].repeat(num_vectors_per_token, 1)
- )
- else:
- for ii in range(len(placeholder_tokens)):
- init_id = tokenizer("a").input_ids[1]
- temp_embedding = embedding_layer.weight[init_id]
- len_emb = temp_embedding.shape[0]
- init_weight = (torch.rand(num_vectors_per_token, len_emb) - 0.5) / 2.0
- initialize_embedding.append(init_weight)
-
- # initialize_embedding = torch.cat(initialize_embedding,dim=0)
-
- token_info_all = []
- for ii in range(len(placeholder_tokens)):
- token_info = tokenizer.get_token_info(placeholder_tokens[ii])
- token_info["embedding"] = initialize_embedding[ii]
- token_info["trainable"] = True
- token_info_all.append(token_info)
- embedding_layer.add_embeddings(token_info_all)
diff --git a/iopaint/model/power_paint/v2/BrushNet_CA.py b/iopaint/model/power_paint/v2/BrushNet_CA.py
new file mode 100644
index 0000000..807f3e6
--- /dev/null
+++ b/iopaint/model/power_paint/v2/BrushNet_CA.py
@@ -0,0 +1,933 @@
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+from torch import nn
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.utils import BaseOutput, logging
+from diffusers.models.attention_processor import (
+ ADDED_KV_ATTENTION_PROCESSORS,
+ CROSS_ATTENTION_PROCESSORS,
+ AttentionProcessor,
+ AttnAddedKVProcessor,
+ AttnProcessor,
+)
+from diffusers.models.embeddings import TextImageProjection, TextImageTimeEmbedding, TextTimeEmbedding, \
+ TimestepEmbedding, Timesteps
+from diffusers.models.modeling_utils import ModelMixin
+from .unet_2d_blocks import (
+ CrossAttnDownBlock2D,
+ DownBlock2D,
+ get_down_block,
+ get_mid_block,
+ get_up_block
+)
+
+from .unet_2d_condition import UNet2DConditionModel
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+@dataclass
+class BrushNetOutput(BaseOutput):
+ """
+ The output of [`BrushNetModel`].
+
+ Args:
+ up_block_res_samples (`tuple[torch.Tensor]`):
+ A tuple of upsample activations at different resolutions for each upsampling block. Each tensor should
+ be of shape `(batch_size, channel * resolution, height //resolution, width // resolution)`. Output can be
+ used to condition the original UNet's upsampling activations.
+ down_block_res_samples (`tuple[torch.Tensor]`):
+ A tuple of downsample activations at different resolutions for each downsampling block. Each tensor should
+ be of shape `(batch_size, channel * resolution, height //resolution, width // resolution)`. Output can be
+ used to condition the original UNet's downsampling activations.
+ mid_down_block_re_sample (`torch.Tensor`):
+ The activation of the midde block (the lowest sample resolution). Each tensor should be of shape
+ `(batch_size, channel * lowest_resolution, height // lowest_resolution, width // lowest_resolution)`.
+ Output can be used to condition the original UNet's middle block activation.
+ """
+
+ up_block_res_samples: Tuple[torch.Tensor]
+ down_block_res_samples: Tuple[torch.Tensor]
+ mid_block_res_sample: torch.Tensor
+
+
+class BrushNetModel(ModelMixin, ConfigMixin):
+ """
+ A BrushNet model.
+
+ Args:
+ in_channels (`int`, defaults to 4):
+ The number of channels in the input sample.
+ flip_sin_to_cos (`bool`, defaults to `True`):
+ Whether to flip the sin to cos in the time embedding.
+ freq_shift (`int`, defaults to 0):
+ The frequency shift to apply to the time embedding.
+ down_block_types (`tuple[str]`, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
+ The tuple of downsample blocks to use.
+ mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`):
+ Block type for middle of UNet, it can be one of `UNetMidBlock2DCrossAttn`, `UNetMidBlock2D`, or
+ `UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped.
+ up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`):
+ The tuple of upsample blocks to use.
+ only_cross_attention (`Union[bool, Tuple[bool]]`, defaults to `False`):
+ block_out_channels (`tuple[int]`, defaults to `(320, 640, 1280, 1280)`):
+ The tuple of output channels for each block.
+ layers_per_block (`int`, defaults to 2):
+ The number of layers per block.
+ downsample_padding (`int`, defaults to 1):
+ The padding to use for the downsampling convolution.
+ mid_block_scale_factor (`float`, defaults to 1):
+ The scale factor to use for the mid block.
+ act_fn (`str`, defaults to "silu"):
+ The activation function to use.
+ norm_num_groups (`int`, *optional*, defaults to 32):
+ The number of groups to use for the normalization. If None, normalization and activation layers is skipped
+ in post-processing.
+ norm_eps (`float`, defaults to 1e-5):
+ The epsilon to use for the normalization.
+ cross_attention_dim (`int`, defaults to 1280):
+ The dimension of the cross attention features.
+ transformer_layers_per_block (`int` or `Tuple[int]`, *optional*, defaults to 1):
+ The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
+ [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+ [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+ encoder_hid_dim (`int`, *optional*, defaults to None):
+ If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
+ dimension to `cross_attention_dim`.
+ encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
+ If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
+ embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
+ attention_head_dim (`Union[int, Tuple[int]]`, defaults to 8):
+ The dimension of the attention heads.
+ use_linear_projection (`bool`, defaults to `False`):
+ class_embed_type (`str`, *optional*, defaults to `None`):
+ The type of class embedding to use which is ultimately summed with the time embeddings. Choose from None,
+ `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
+ addition_embed_type (`str`, *optional*, defaults to `None`):
+ Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
+ "text". "text" will use the `TextTimeEmbedding` layer.
+ num_class_embeds (`int`, *optional*, defaults to 0):
+ Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
+ class conditioning with `class_embed_type` equal to `None`.
+ upcast_attention (`bool`, defaults to `False`):
+ resnet_time_scale_shift (`str`, defaults to `"default"`):
+ Time scale shift config for ResNet blocks (see `ResnetBlock2D`). Choose from `default` or `scale_shift`.
+ projection_class_embeddings_input_dim (`int`, *optional*, defaults to `None`):
+ The dimension of the `class_labels` input when `class_embed_type="projection"`. Required when
+ `class_embed_type="projection"`.
+ brushnet_conditioning_channel_order (`str`, defaults to `"rgb"`):
+ The channel order of conditional image. Will convert to `rgb` if it's `bgr`.
+ conditioning_embedding_out_channels (`tuple[int]`, *optional*, defaults to `(16, 32, 96, 256)`):
+ The tuple of output channel for each block in the `conditioning_embedding` layer.
+ global_pool_conditions (`bool`, defaults to `False`):
+ TODO(Patrick) - unused parameter.
+ addition_embed_type_num_heads (`int`, defaults to 64):
+ The number of heads to use for the `TextTimeEmbedding` layer.
+ """
+
+ _supports_gradient_checkpointing = True
+
+ @register_to_config
+ def __init__(
+ self,
+ in_channels: int = 4,
+ conditioning_channels: int = 5,
+ flip_sin_to_cos: bool = True,
+ freq_shift: int = 0,
+ down_block_types: Tuple[str, ...] = (
+ "CrossAttnDownBlock2D",
+ "CrossAttnDownBlock2D",
+ "CrossAttnDownBlock2D",
+ "DownBlock2D",
+ ),
+ mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
+ up_block_types: Tuple[str, ...] = (
+ "UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"
+ ),
+ only_cross_attention: Union[bool, Tuple[bool]] = False,
+ block_out_channels: Tuple[int, ...] = (320, 640, 1280, 1280),
+ layers_per_block: int = 2,
+ downsample_padding: int = 1,
+ mid_block_scale_factor: float = 1,
+ act_fn: str = "silu",
+ norm_num_groups: Optional[int] = 32,
+ norm_eps: float = 1e-5,
+ cross_attention_dim: int = 1280,
+ transformer_layers_per_block: Union[int, Tuple[int, ...]] = 1,
+ encoder_hid_dim: Optional[int] = None,
+ encoder_hid_dim_type: Optional[str] = None,
+ attention_head_dim: Union[int, Tuple[int, ...]] = 8,
+ num_attention_heads: Optional[Union[int, Tuple[int, ...]]] = None,
+ use_linear_projection: bool = False,
+ class_embed_type: Optional[str] = None,
+ addition_embed_type: Optional[str] = None,
+ addition_time_embed_dim: Optional[int] = None,
+ num_class_embeds: Optional[int] = None,
+ upcast_attention: bool = False,
+ resnet_time_scale_shift: str = "default",
+ projection_class_embeddings_input_dim: Optional[int] = None,
+ brushnet_conditioning_channel_order: str = "rgb",
+ conditioning_embedding_out_channels: Optional[Tuple[int, ...]] = (16, 32, 96, 256),
+ global_pool_conditions: bool = False,
+ addition_embed_type_num_heads: int = 64,
+ ):
+ super().__init__()
+
+ # If `num_attention_heads` is not defined (which is the case for most models)
+ # it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
+ # The reason for this behavior is to correct for incorrectly named variables that were introduced
+ # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
+ # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
+ # which is why we correct for the naming here.
+ num_attention_heads = num_attention_heads or attention_head_dim
+
+ # Check inputs
+ if len(down_block_types) != len(up_block_types):
+ raise ValueError(
+ f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
+ )
+
+ if len(block_out_channels) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
+ )
+
+ if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
+ )
+
+ if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
+ )
+
+ if isinstance(transformer_layers_per_block, int):
+ transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types)
+
+ # input
+ conv_in_kernel = 3
+ conv_in_padding = (conv_in_kernel - 1) // 2
+ self.conv_in_condition = nn.Conv2d(
+ in_channels + conditioning_channels, block_out_channels[0], kernel_size=conv_in_kernel,
+ padding=conv_in_padding
+ )
+
+ # time
+ time_embed_dim = block_out_channels[0] * 4
+ self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
+ timestep_input_dim = block_out_channels[0]
+ self.time_embedding = TimestepEmbedding(
+ timestep_input_dim,
+ time_embed_dim,
+ act_fn=act_fn,
+ )
+
+ if encoder_hid_dim_type is None and encoder_hid_dim is not None:
+ encoder_hid_dim_type = "text_proj"
+ self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
+ logger.info("encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined.")
+
+ if encoder_hid_dim is None and encoder_hid_dim_type is not None:
+ raise ValueError(
+ f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}."
+ )
+
+ if encoder_hid_dim_type == "text_proj":
+ self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim)
+ elif encoder_hid_dim_type == "text_image_proj":
+ # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+ # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+ # case when `addition_embed_type == "text_image_proj"` (Kadinsky 2.1)`
+ self.encoder_hid_proj = TextImageProjection(
+ text_embed_dim=encoder_hid_dim,
+ image_embed_dim=cross_attention_dim,
+ cross_attention_dim=cross_attention_dim,
+ )
+
+ elif encoder_hid_dim_type is not None:
+ raise ValueError(
+ f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'."
+ )
+ else:
+ self.encoder_hid_proj = None
+
+ # class embedding
+ if class_embed_type is None and num_class_embeds is not None:
+ self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
+ elif class_embed_type == "timestep":
+ self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+ elif class_embed_type == "identity":
+ self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
+ elif class_embed_type == "projection":
+ if projection_class_embeddings_input_dim is None:
+ raise ValueError(
+ "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
+ )
+ # The projection `class_embed_type` is the same as the timestep `class_embed_type` except
+ # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
+ # 2. it projects from an arbitrary input dimension.
+ #
+ # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
+ # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
+ # As a result, `TimestepEmbedding` can be passed arbitrary vectors.
+ self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+ else:
+ self.class_embedding = None
+
+ if addition_embed_type == "text":
+ if encoder_hid_dim is not None:
+ text_time_embedding_from_dim = encoder_hid_dim
+ else:
+ text_time_embedding_from_dim = cross_attention_dim
+
+ self.add_embedding = TextTimeEmbedding(
+ text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads
+ )
+ elif addition_embed_type == "text_image":
+ # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+ # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+ # case when `addition_embed_type == "text_image"` (Kadinsky 2.1)`
+ self.add_embedding = TextImageTimeEmbedding(
+ text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim
+ )
+ elif addition_embed_type == "text_time":
+ self.add_time_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift)
+ self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+
+ elif addition_embed_type is not None:
+ raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.")
+
+ self.down_blocks = nn.ModuleList([])
+ self.brushnet_down_blocks = nn.ModuleList([])
+
+ if isinstance(only_cross_attention, bool):
+ only_cross_attention = [only_cross_attention] * len(down_block_types)
+
+ if isinstance(attention_head_dim, int):
+ attention_head_dim = (attention_head_dim,) * len(down_block_types)
+
+ if isinstance(num_attention_heads, int):
+ num_attention_heads = (num_attention_heads,) * len(down_block_types)
+
+ # down
+ output_channel = block_out_channels[0]
+
+ brushnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
+ brushnet_block = zero_module(brushnet_block)
+ self.brushnet_down_blocks.append(brushnet_block)
+
+ for i, down_block_type in enumerate(down_block_types):
+ input_channel = output_channel
+ output_channel = block_out_channels[i]
+ is_final_block = i == len(block_out_channels) - 1
+
+ down_block = get_down_block(
+ down_block_type,
+ num_layers=layers_per_block,
+ transformer_layers_per_block=transformer_layers_per_block[i],
+ in_channels=input_channel,
+ out_channels=output_channel,
+ temb_channels=time_embed_dim,
+ add_downsample=not is_final_block,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ resnet_groups=norm_num_groups,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=num_attention_heads[i],
+ attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
+ downsample_padding=downsample_padding,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention[i],
+ upcast_attention=upcast_attention,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ )
+ self.down_blocks.append(down_block)
+
+ for _ in range(layers_per_block):
+ brushnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
+ brushnet_block = zero_module(brushnet_block)
+ self.brushnet_down_blocks.append(brushnet_block)
+
+ if not is_final_block:
+ brushnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
+ brushnet_block = zero_module(brushnet_block)
+ self.brushnet_down_blocks.append(brushnet_block)
+
+ # mid
+ mid_block_channel = block_out_channels[-1]
+
+ brushnet_block = nn.Conv2d(mid_block_channel, mid_block_channel, kernel_size=1)
+ brushnet_block = zero_module(brushnet_block)
+ self.brushnet_mid_block = brushnet_block
+
+ self.mid_block = get_mid_block(
+ mid_block_type,
+ transformer_layers_per_block=transformer_layers_per_block[-1],
+ in_channels=mid_block_channel,
+ temb_channels=time_embed_dim,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ output_scale_factor=mid_block_scale_factor,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=num_attention_heads[-1],
+ resnet_groups=norm_num_groups,
+ use_linear_projection=use_linear_projection,
+ upcast_attention=upcast_attention,
+ )
+
+ # count how many layers upsample the images
+ self.num_upsamplers = 0
+
+ # up
+ reversed_block_out_channels = list(reversed(block_out_channels))
+ reversed_num_attention_heads = list(reversed(num_attention_heads))
+ reversed_transformer_layers_per_block = (list(reversed(transformer_layers_per_block)))
+ only_cross_attention = list(reversed(only_cross_attention))
+
+ output_channel = reversed_block_out_channels[0]
+
+ self.up_blocks = nn.ModuleList([])
+ self.brushnet_up_blocks = nn.ModuleList([])
+
+ for i, up_block_type in enumerate(up_block_types):
+ is_final_block = i == len(block_out_channels) - 1
+
+ prev_output_channel = output_channel
+ output_channel = reversed_block_out_channels[i]
+ input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
+
+ # add upsample block for all BUT final layer
+ if not is_final_block:
+ add_upsample = True
+ self.num_upsamplers += 1
+ else:
+ add_upsample = False
+
+ up_block = get_up_block(
+ up_block_type,
+ num_layers=layers_per_block + 1,
+ transformer_layers_per_block=reversed_transformer_layers_per_block[i],
+ in_channels=input_channel,
+ out_channels=output_channel,
+ prev_output_channel=prev_output_channel,
+ temb_channels=time_embed_dim,
+ add_upsample=add_upsample,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ resolution_idx=i,
+ resnet_groups=norm_num_groups,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=reversed_num_attention_heads[i],
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention[i],
+ upcast_attention=upcast_attention,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
+ )
+ self.up_blocks.append(up_block)
+ prev_output_channel = output_channel
+
+ for _ in range(layers_per_block + 1):
+ brushnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
+ brushnet_block = zero_module(brushnet_block)
+ self.brushnet_up_blocks.append(brushnet_block)
+
+ if not is_final_block:
+ brushnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
+ brushnet_block = zero_module(brushnet_block)
+ self.brushnet_up_blocks.append(brushnet_block)
+
+ @classmethod
+ def from_unet(
+ cls,
+ unet: UNet2DConditionModel,
+ brushnet_conditioning_channel_order: str = "rgb",
+ conditioning_embedding_out_channels: Optional[Tuple[int, ...]] = (16, 32, 96, 256),
+ load_weights_from_unet: bool = True,
+ conditioning_channels: int = 5,
+ ):
+ r"""
+ Instantiate a [`BrushNetModel`] from [`UNet2DConditionModel`].
+
+ Parameters:
+ unet (`UNet2DConditionModel`):
+ The UNet model weights to copy to the [`BrushNetModel`]. All configuration options are also copied
+ where applicable.
+ """
+ transformer_layers_per_block = (
+ unet.config.transformer_layers_per_block if "transformer_layers_per_block" in unet.config else 1
+ )
+ encoder_hid_dim = unet.config.encoder_hid_dim if "encoder_hid_dim" in unet.config else None
+ encoder_hid_dim_type = unet.config.encoder_hid_dim_type if "encoder_hid_dim_type" in unet.config else None
+ addition_embed_type = unet.config.addition_embed_type if "addition_embed_type" in unet.config else None
+ addition_time_embed_dim = (
+ unet.config.addition_time_embed_dim if "addition_time_embed_dim" in unet.config else None
+ )
+
+ brushnet = cls(
+ in_channels=unet.config.in_channels,
+ conditioning_channels=conditioning_channels,
+ flip_sin_to_cos=unet.config.flip_sin_to_cos,
+ freq_shift=unet.config.freq_shift,
+ # down_block_types=['DownBlock2D','DownBlock2D','DownBlock2D','DownBlock2D'],
+ down_block_types=["CrossAttnDownBlock2D",
+ "CrossAttnDownBlock2D",
+ "CrossAttnDownBlock2D",
+ "DownBlock2D", ],
+ # mid_block_type='MidBlock2D',
+ mid_block_type="UNetMidBlock2DCrossAttn",
+ # up_block_types=['UpBlock2D','UpBlock2D','UpBlock2D','UpBlock2D'],
+ up_block_types=["UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"],
+ only_cross_attention=unet.config.only_cross_attention,
+ block_out_channels=unet.config.block_out_channels,
+ layers_per_block=unet.config.layers_per_block,
+ downsample_padding=unet.config.downsample_padding,
+ mid_block_scale_factor=unet.config.mid_block_scale_factor,
+ act_fn=unet.config.act_fn,
+ norm_num_groups=unet.config.norm_num_groups,
+ norm_eps=unet.config.norm_eps,
+ cross_attention_dim=unet.config.cross_attention_dim,
+ transformer_layers_per_block=transformer_layers_per_block,
+ encoder_hid_dim=encoder_hid_dim,
+ encoder_hid_dim_type=encoder_hid_dim_type,
+ attention_head_dim=unet.config.attention_head_dim,
+ num_attention_heads=unet.config.num_attention_heads,
+ use_linear_projection=unet.config.use_linear_projection,
+ class_embed_type=unet.config.class_embed_type,
+ addition_embed_type=addition_embed_type,
+ addition_time_embed_dim=addition_time_embed_dim,
+ num_class_embeds=unet.config.num_class_embeds,
+ upcast_attention=unet.config.upcast_attention,
+ resnet_time_scale_shift=unet.config.resnet_time_scale_shift,
+ projection_class_embeddings_input_dim=unet.config.projection_class_embeddings_input_dim,
+ brushnet_conditioning_channel_order=brushnet_conditioning_channel_order,
+ conditioning_embedding_out_channels=conditioning_embedding_out_channels,
+ )
+
+ if load_weights_from_unet:
+ conv_in_condition_weight = torch.zeros_like(brushnet.conv_in_condition.weight)
+ conv_in_condition_weight[:, :4, ...] = unet.conv_in.weight
+ conv_in_condition_weight[:, 4:8, ...] = unet.conv_in.weight
+ brushnet.conv_in_condition.weight = torch.nn.Parameter(conv_in_condition_weight)
+ brushnet.conv_in_condition.bias = unet.conv_in.bias
+
+ brushnet.time_proj.load_state_dict(unet.time_proj.state_dict())
+ brushnet.time_embedding.load_state_dict(unet.time_embedding.state_dict())
+
+ if brushnet.class_embedding:
+ brushnet.class_embedding.load_state_dict(unet.class_embedding.state_dict())
+
+ brushnet.down_blocks.load_state_dict(unet.down_blocks.state_dict(), strict=False)
+ brushnet.mid_block.load_state_dict(unet.mid_block.state_dict(), strict=False)
+ brushnet.up_blocks.load_state_dict(unet.up_blocks.state_dict(), strict=False)
+
+ return brushnet.to(unet.dtype)
+
+ @property
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+ def attn_processors(self) -> Dict[str, AttentionProcessor]:
+ r"""
+ Returns:
+ `dict` of attention processors: A dictionary containing all attention processors used in the model with
+ indexed by its weight name.
+ """
+ # set recursively
+ processors = {}
+
+ def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+ if hasattr(module, "get_processor"):
+ processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
+
+ for sub_name, child in module.named_children():
+ fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+ return processors
+
+ for name, module in self.named_children():
+ fn_recursive_add_processors(name, module, processors)
+
+ return processors
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+ def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+ r"""
+ Sets the attention processor to use to compute attention.
+
+ Parameters:
+ processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+ The instantiated processor class or a dictionary of processor classes that will be set as the processor
+ for **all** `Attention` layers.
+
+ If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+ processor. This is strongly recommended when setting trainable attention processors.
+
+ """
+ count = len(self.attn_processors.keys())
+
+ if isinstance(processor, dict) and len(processor) != count:
+ raise ValueError(
+ f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+ f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+ )
+
+ def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+ if hasattr(module, "set_processor"):
+ if not isinstance(processor, dict):
+ module.set_processor(processor)
+ else:
+ module.set_processor(processor.pop(f"{name}.processor"))
+
+ for sub_name, child in module.named_children():
+ fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+ for name, module in self.named_children():
+ fn_recursive_attn_processor(name, module, processor)
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
+ def set_default_attn_processor(self):
+ """
+ Disables custom attention processors and sets the default attention implementation.
+ """
+ if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+ processor = AttnAddedKVProcessor()
+ elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+ processor = AttnProcessor()
+ else:
+ raise ValueError(
+ f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+ )
+
+ self.set_attn_processor(processor)
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attention_slice
+ def set_attention_slice(self, slice_size: Union[str, int, List[int]]) -> None:
+ r"""
+ Enable sliced attention computation.
+
+ When this option is enabled, the attention module splits the input tensor in slices to compute attention in
+ several steps. This is useful for saving some memory in exchange for a small decrease in speed.
+
+ Args:
+ slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+ When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
+ `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
+ provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+ must be a multiple of `slice_size`.
+ """
+ sliceable_head_dims = []
+
+ def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
+ if hasattr(module, "set_attention_slice"):
+ sliceable_head_dims.append(module.sliceable_head_dim)
+
+ for child in module.children():
+ fn_recursive_retrieve_sliceable_dims(child)
+
+ # retrieve number of attention layers
+ for module in self.children():
+ fn_recursive_retrieve_sliceable_dims(module)
+
+ num_sliceable_layers = len(sliceable_head_dims)
+
+ if slice_size == "auto":
+ # half the attention head size is usually a good trade-off between
+ # speed and memory
+ slice_size = [dim // 2 for dim in sliceable_head_dims]
+ elif slice_size == "max":
+ # make smallest slice possible
+ slice_size = num_sliceable_layers * [1]
+
+ slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size
+
+ if len(slice_size) != len(sliceable_head_dims):
+ raise ValueError(
+ f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
+ f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
+ )
+
+ for i in range(len(slice_size)):
+ size = slice_size[i]
+ dim = sliceable_head_dims[i]
+ if size is not None and size > dim:
+ raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
+
+ # Recursively walk through all the children.
+ # Any children which exposes the set_attention_slice method
+ # gets the message
+ def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]):
+ if hasattr(module, "set_attention_slice"):
+ module.set_attention_slice(slice_size.pop())
+
+ for child in module.children():
+ fn_recursive_set_attention_slice(child, slice_size)
+
+ reversed_slice_size = list(reversed(slice_size))
+ for module in self.children():
+ fn_recursive_set_attention_slice(module, reversed_slice_size)
+
+ def _set_gradient_checkpointing(self, module, value: bool = False) -> None:
+ if isinstance(module, (CrossAttnDownBlock2D, DownBlock2D)):
+ module.gradient_checkpointing = value
+
+ def forward(
+ self,
+ sample: torch.FloatTensor,
+ timestep: Union[torch.Tensor, float, int],
+ encoder_hidden_states: torch.Tensor,
+ brushnet_cond: torch.FloatTensor,
+ conditioning_scale: float = 1.0,
+ class_labels: Optional[torch.Tensor] = None,
+ timestep_cond: Optional[torch.Tensor] = None,
+ attention_mask: Optional[torch.Tensor] = None,
+ added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ guess_mode: bool = False,
+ return_dict: bool = True,
+ ) -> Union[BrushNetOutput, Tuple[Tuple[torch.FloatTensor, ...], torch.FloatTensor]]:
+ """
+ The [`BrushNetModel`] forward method.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The noisy input tensor.
+ timestep (`Union[torch.Tensor, float, int]`):
+ The number of timesteps to denoise an input.
+ encoder_hidden_states (`torch.Tensor`):
+ The encoder hidden states.
+ brushnet_cond (`torch.FloatTensor`):
+ The conditional input tensor of shape `(batch_size, sequence_length, hidden_size)`.
+ conditioning_scale (`float`, defaults to `1.0`):
+ The scale factor for BrushNet outputs.
+ class_labels (`torch.Tensor`, *optional*, defaults to `None`):
+ Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
+ timestep_cond (`torch.Tensor`, *optional*, defaults to `None`):
+ Additional conditional embeddings for timestep. If provided, the embeddings will be summed with the
+ timestep_embedding passed through the `self.time_embedding` layer to obtain the final timestep
+ embeddings.
+ attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
+ An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+ is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+ negative values to the attention scores corresponding to "discard" tokens.
+ added_cond_kwargs (`dict`):
+ Additional conditions for the Stable Diffusion XL UNet.
+ cross_attention_kwargs (`dict[str]`, *optional*, defaults to `None`):
+ A kwargs dictionary that if specified is passed along to the `AttnProcessor`.
+ guess_mode (`bool`, defaults to `False`):
+ In this mode, the BrushNet encoder tries its best to recognize the input content of the input even if
+ you remove all prompts. A `guidance_scale` between 3.0 and 5.0 is recommended.
+ return_dict (`bool`, defaults to `True`):
+ Whether or not to return a [`~models.brushnet.BrushNetOutput`] instead of a plain tuple.
+
+ Returns:
+ [`~models.brushnet.BrushNetOutput`] **or** `tuple`:
+ If `return_dict` is `True`, a [`~models.brushnet.BrushNetOutput`] is returned, otherwise a tuple is
+ returned where the first element is the sample tensor.
+ """
+ # check channel order
+ channel_order = self.config.brushnet_conditioning_channel_order
+
+ if channel_order == "rgb":
+ # in rgb order by default
+ ...
+ elif channel_order == "bgr":
+ brushnet_cond = torch.flip(brushnet_cond, dims=[1])
+ else:
+ raise ValueError(f"unknown `brushnet_conditioning_channel_order`: {channel_order}")
+
+ # prepare attention_mask
+ if attention_mask is not None:
+ attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+ attention_mask = attention_mask.unsqueeze(1)
+
+ # 1. time
+ timesteps = timestep
+ if not torch.is_tensor(timesteps):
+ # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+ # This would be a good case for the `match` statement (Python 3.10+)
+ is_mps = sample.device.type == "mps"
+ if isinstance(timestep, float):
+ dtype = torch.float32 if is_mps else torch.float64
+ else:
+ dtype = torch.int32 if is_mps else torch.int64
+ timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+ elif len(timesteps.shape) == 0:
+ timesteps = timesteps[None].to(sample.device)
+
+ # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+ timesteps = timesteps.expand(sample.shape[0])
+
+ t_emb = self.time_proj(timesteps)
+
+ # timesteps does not contain any weights and will always return f32 tensors
+ # but time_embedding might actually be running in fp16. so we need to cast here.
+ # there might be better ways to encapsulate this.
+ t_emb = t_emb.to(dtype=sample.dtype)
+
+ emb = self.time_embedding(t_emb, timestep_cond)
+ aug_emb = None
+
+ if self.class_embedding is not None:
+ if class_labels is None:
+ raise ValueError("class_labels should be provided when num_class_embeds > 0")
+
+ if self.config.class_embed_type == "timestep":
+ class_labels = self.time_proj(class_labels)
+
+ class_emb = self.class_embedding(class_labels).to(dtype=self.dtype)
+ emb = emb + class_emb
+
+ if self.config.addition_embed_type is not None:
+ if self.config.addition_embed_type == "text":
+ aug_emb = self.add_embedding(encoder_hidden_states)
+
+ elif self.config.addition_embed_type == "text_time":
+ if "text_embeds" not in added_cond_kwargs:
+ raise ValueError(
+ f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
+ )
+ text_embeds = added_cond_kwargs.get("text_embeds")
+ if "time_ids" not in added_cond_kwargs:
+ raise ValueError(
+ f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
+ )
+ time_ids = added_cond_kwargs.get("time_ids")
+ time_embeds = self.add_time_proj(time_ids.flatten())
+ time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
+
+ add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
+ add_embeds = add_embeds.to(emb.dtype)
+ aug_emb = self.add_embedding(add_embeds)
+
+ emb = emb + aug_emb if aug_emb is not None else emb
+
+ # 2. pre-process
+ brushnet_cond = torch.concat([sample, brushnet_cond], 1)
+ sample = self.conv_in_condition(brushnet_cond)
+
+ # 3. down
+ down_block_res_samples = (sample,)
+ for downsample_block in self.down_blocks:
+ if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
+ sample, res_samples = downsample_block(
+ hidden_states=sample,
+ temb=emb,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=attention_mask,
+ cross_attention_kwargs=cross_attention_kwargs,
+ )
+ else:
+ sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
+
+ down_block_res_samples += res_samples
+
+ # 4. PaintingNet down blocks
+ brushnet_down_block_res_samples = ()
+ for down_block_res_sample, brushnet_down_block in zip(down_block_res_samples, self.brushnet_down_blocks):
+ down_block_res_sample = brushnet_down_block(down_block_res_sample)
+ brushnet_down_block_res_samples = brushnet_down_block_res_samples + (down_block_res_sample,)
+
+ # 5. mid
+ if self.mid_block is not None:
+ if hasattr(self.mid_block, "has_cross_attention") and self.mid_block.has_cross_attention:
+ sample = self.mid_block(
+ sample,
+ emb,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=attention_mask,
+ cross_attention_kwargs=cross_attention_kwargs,
+ )
+ else:
+ sample = self.mid_block(sample, emb)
+
+ # 6. BrushNet mid blocks
+ brushnet_mid_block_res_sample = self.brushnet_mid_block(sample)
+
+ # 7. up
+ up_block_res_samples = ()
+ for i, upsample_block in enumerate(self.up_blocks):
+ is_final_block = i == len(self.up_blocks) - 1
+
+ res_samples = down_block_res_samples[-len(upsample_block.resnets):]
+ down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
+
+ # if we have not reached the final block and need to forward the
+ # upsample size, we do it here
+ if not is_final_block:
+ upsample_size = down_block_res_samples[-1].shape[2:]
+
+ if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
+ sample, up_res_samples = upsample_block(
+ hidden_states=sample,
+ temb=emb,
+ res_hidden_states_tuple=res_samples,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ upsample_size=upsample_size,
+ attention_mask=attention_mask,
+ return_res_samples=True
+ )
+ else:
+ sample, up_res_samples = upsample_block(
+ hidden_states=sample,
+ temb=emb,
+ res_hidden_states_tuple=res_samples,
+ upsample_size=upsample_size,
+ return_res_samples=True
+ )
+
+ up_block_res_samples += up_res_samples
+
+ # 8. BrushNet up blocks
+ brushnet_up_block_res_samples = ()
+ for up_block_res_sample, brushnet_up_block in zip(up_block_res_samples, self.brushnet_up_blocks):
+ up_block_res_sample = brushnet_up_block(up_block_res_sample)
+ brushnet_up_block_res_samples = brushnet_up_block_res_samples + (up_block_res_sample,)
+
+ # 6. scaling
+ if guess_mode and not self.config.global_pool_conditions:
+ scales = torch.logspace(-1, 0,
+ len(brushnet_down_block_res_samples) + 1 + len(brushnet_up_block_res_samples),
+ device=sample.device) # 0.1 to 1.0
+ scales = scales * conditioning_scale
+
+ brushnet_down_block_res_samples = [sample * scale for sample, scale in zip(brushnet_down_block_res_samples,
+ scales[:len(
+ brushnet_down_block_res_samples)])]
+ brushnet_mid_block_res_sample = brushnet_mid_block_res_sample * scales[len(brushnet_down_block_res_samples)]
+ brushnet_up_block_res_samples = [sample * scale for sample, scale in zip(brushnet_up_block_res_samples,
+ scales[
+ len(brushnet_down_block_res_samples) + 1:])]
+ else:
+ brushnet_down_block_res_samples = [sample * conditioning_scale for sample in
+ brushnet_down_block_res_samples]
+ brushnet_mid_block_res_sample = brushnet_mid_block_res_sample * conditioning_scale
+ brushnet_up_block_res_samples = [sample * conditioning_scale for sample in brushnet_up_block_res_samples]
+
+ if self.config.global_pool_conditions:
+ brushnet_down_block_res_samples = [
+ torch.mean(sample, dim=(2, 3), keepdim=True) for sample in brushnet_down_block_res_samples
+ ]
+ brushnet_mid_block_res_sample = torch.mean(brushnet_mid_block_res_sample, dim=(2, 3), keepdim=True)
+ brushnet_up_block_res_samples = [
+ torch.mean(sample, dim=(2, 3), keepdim=True) for sample in brushnet_up_block_res_samples
+ ]
+
+ if not return_dict:
+ return (brushnet_down_block_res_samples, brushnet_mid_block_res_sample, brushnet_up_block_res_samples)
+
+ return BrushNetOutput(
+ down_block_res_samples=brushnet_down_block_res_samples,
+ mid_block_res_sample=brushnet_mid_block_res_sample,
+ up_block_res_samples=brushnet_up_block_res_samples
+ )
+
+
+def zero_module(module):
+ for p in module.parameters():
+ nn.init.zeros_(p)
+ return module
diff --git a/iopaint/model/power_paint/v2/pipeline_PowerPaint_Brushnet_CA.py b/iopaint/model/power_paint/v2/pipeline_PowerPaint_Brushnet_CA.py
new file mode 100644
index 0000000..936f383
--- /dev/null
+++ b/iopaint/model/power_paint/v2/pipeline_PowerPaint_Brushnet_CA.py
@@ -0,0 +1,1494 @@
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+import torch.nn.functional as F
+from diffusers import StableDiffusionMixin
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
+from diffusers.loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from diffusers.models import AutoencoderKL, ImageProjection
+from diffusers.models.lora import adjust_lora_scale_text_encoder
+from diffusers.schedulers import KarrasDiffusionSchedulers
+from diffusers.utils import (
+ USE_PEFT_BACKEND,
+ deprecate,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from diffusers.utils.torch_utils import is_compiled_module, is_torch_version, randn_tensor
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.pipelines.stable_diffusion.pipeline_output import StableDiffusionPipelineOutput
+from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+
+from .BrushNet_CA import BrushNetModel
+from .unet_2d_condition import UNet2DConditionModel
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ from diffusers import StableDiffusionBrushNetPipeline, BrushNetModel, UniPCMultistepScheduler
+ from diffusers.utils import load_image
+ import torch
+ import cv2
+ import numpy as np
+ from PIL import Image
+
+ base_model_path = "runwayml/stable-diffusion-v1-5"
+ brushnet_path = "ckpt_path"
+
+ brushnet = BrushNetModel.from_pretrained(brushnet_path, torch_dtype=torch.float16)
+ pipe = StableDiffusionBrushNetPipeline.from_pretrained(
+ base_model_path, brushnet=brushnet, torch_dtype=torch.float16, low_cpu_mem_usage=False
+ )
+
+ # speed up diffusion process with faster scheduler and memory optimization
+ pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
+ # remove following line if xformers is not installed or when using Torch 2.0.
+ # pipe.enable_xformers_memory_efficient_attention()
+ # memory optimization.
+ pipe.enable_model_cpu_offload()
+
+ image_path="examples/brushnet/src/test_image.jpg"
+ mask_path="examples/brushnet/src/test_mask.jpg"
+ caption="A cake on the table."
+
+ init_image = cv2.imread(image_path)
+ mask_image = 1.*(cv2.imread(mask_path).sum(-1)>255)[:,:,np.newaxis]
+ init_image = init_image * (1-mask_image)
+
+ init_image = Image.fromarray(init_image.astype(np.uint8)).convert("RGB")
+ mask_image = Image.fromarray(mask_image.astype(np.uint8).repeat(3,-1)*255).convert("RGB")
+
+ generator = torch.Generator("cuda").manual_seed(1234)
+
+ image = pipe(
+ caption,
+ init_image,
+ mask_image,
+ num_inference_steps=50,
+ generator=generator,
+ paintingnet_conditioning_scale=1.0
+ ).images[0]
+ image.save("output.png")
+ ```
+"""
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+ scheduler,
+ num_inference_steps: Optional[int] = None,
+ device: Optional[Union[str, torch.device]] = None,
+ timesteps: Optional[List[int]] = None,
+ **kwargs,
+):
+ """
+ Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+ custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+ Args:
+ scheduler (`SchedulerMixin`):
+ The scheduler to get timesteps from.
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model. If used,
+ `timesteps` must be `None`.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
+ timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps`
+ must be `None`.
+
+ Returns:
+ `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+ second element is the number of inference steps.
+ """
+ if timesteps is not None:
+ accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+ if not accepts_timesteps:
+ raise ValueError(
+ f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+ f" timestep schedules. Please check whether you are using the correct scheduler."
+ )
+ scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ num_inference_steps = len(timesteps)
+ else:
+ scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ return timesteps, num_inference_steps
+
+
+class StableDiffusionPowerPaintBrushNetPipeline(
+ DiffusionPipeline,
+ StableDiffusionMixin,
+ TextualInversionLoaderMixin,
+ LoraLoaderMixin,
+ IPAdapterMixin,
+ FromSingleFileMixin,
+):
+ r"""
+ Pipeline for text-to-image generation using Stable Diffusion with BrushNet guidance.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+ - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+ - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ brushnet ([`BrushNetModel`]`):
+ Provides additional conditioning to the `unet` during the denoising process.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ """
+
+ model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae"
+ _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
+ _exclude_from_cpu_offload = ["safety_checker"]
+ _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ text_encoder_brushnet: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ brushnet: BrushNetModel,
+ scheduler: KarrasDiffusionSchedulers,
+ safety_checker: StableDiffusionSafetyChecker,
+ feature_extractor: CLIPImageProcessor,
+ image_encoder: CLIPVisionModelWithProjection = None,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ text_encoder_brushnet=text_encoder_brushnet,
+ tokenizer=tokenizer,
+ unet=unet,
+ brushnet=brushnet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ image_encoder=image_encoder,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True)
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ promptA,
+ promptB,
+ t,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_promptA=None,
+ negative_promptB=None,
+ t_nag=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ prompt = promptA
+ negative_prompt = negative_promptA
+
+ if promptA is not None and isinstance(promptA, str):
+ batch_size = 1
+ elif promptA is not None and isinstance(promptA, list):
+ batch_size = len(promptA)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: procecss multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ promptA = self.maybe_convert_prompt(promptA, self.tokenizer)
+
+ text_inputsA = self.tokenizer(
+ promptA,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_inputsB = self.tokenizer(
+ promptB,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_idsA = text_inputsA.input_ids
+ text_input_idsB = text_inputsB.input_ids
+ untruncated_ids = self.tokenizer(promptA, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_idsA.shape[-1] and not torch.equal(
+ text_input_idsA, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1: -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder_brushnet.config,
+ "use_attention_mask") and self.text_encoder_brushnet.config.use_attention_mask:
+ attention_mask = text_inputsA.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ # print("text_input_idsA: ",text_input_idsA)
+ # print("text_input_idsB: ",text_input_idsB)
+ # print('t: ',t)
+
+ prompt_embedsA = self.text_encoder_brushnet(
+ text_input_idsA.to(device),
+ attention_mask=attention_mask,
+ )
+ prompt_embedsA = prompt_embedsA[0]
+
+ prompt_embedsB = self.text_encoder_brushnet(
+ text_input_idsB.to(device),
+ attention_mask=attention_mask,
+ )
+ prompt_embedsB = prompt_embedsB[0]
+ prompt_embeds = prompt_embedsA * (t) + (1 - t) * prompt_embedsB
+ # print("prompt_embeds: ",prompt_embeds)
+
+ if self.text_encoder_brushnet is not None:
+ prompt_embeds_dtype = self.text_encoder_brushnet.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokensA: List[str]
+ uncond_tokensB: List[str]
+ if negative_prompt is None:
+ uncond_tokensA = [""] * batch_size
+ uncond_tokensB = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokensA = [negative_promptA]
+ uncond_tokensB = [negative_promptB]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokensA = negative_promptA
+ uncond_tokensB = negative_promptB
+
+ # textual inversion: procecss multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokensA = self.maybe_convert_prompt(uncond_tokensA, self.tokenizer)
+ uncond_tokensB = self.maybe_convert_prompt(uncond_tokensB, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_inputA = self.tokenizer(
+ uncond_tokensA,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ uncond_inputB = self.tokenizer(
+ uncond_tokensB,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder_brushnet.config,
+ "use_attention_mask") and self.text_encoder_brushnet.config.use_attention_mask:
+ attention_mask = uncond_inputA.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embedsA = self.text_encoder_brushnet(
+ uncond_inputA.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embedsB = self.text_encoder_brushnet(
+ uncond_inputB.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embedsA[0] * (t_nag) + (1 - t_nag) * negative_prompt_embedsB[0]
+
+ # negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ # print("prompt_embeds: ",prompt_embeds)
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ # print('1 ',prompt,negative_prompt)
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+ # print('2 ',prompt,negative_prompt)
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+ # print('3 ',prompt,negative_prompt)
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ # print('4 ',prompt,negative_prompt)
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ # print('5 ',prompt,negative_prompt)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ # print(prompt, text_input_ids)
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1: -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ # print("neg: ", uncond_input.input_ids)
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+ def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ if output_hidden_states:
+ image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+ image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_enc_hidden_states = self.image_encoder(
+ torch.zeros_like(image), output_hidden_states=True
+ ).hidden_states[-2]
+ uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+ num_images_per_prompt, dim=0
+ )
+ return image_enc_hidden_states, uncond_image_enc_hidden_states
+ else:
+ image_embeds = self.image_encoder(image).image_embeds
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_embeds = torch.zeros_like(image_embeds)
+
+ return image_embeds, uncond_image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
+ def prepare_ip_adapter_image_embeds(
+ self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
+ ):
+ if ip_adapter_image_embeds is None:
+ if not isinstance(ip_adapter_image, list):
+ ip_adapter_image = [ip_adapter_image]
+
+ if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
+ raise ValueError(
+ f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+ )
+
+ image_embeds = []
+ for single_ip_adapter_image, image_proj_layer in zip(
+ ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
+ ):
+ output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+ single_image_embeds, single_negative_image_embeds = self.encode_image(
+ single_ip_adapter_image, device, 1, output_hidden_state
+ )
+ single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
+ single_negative_image_embeds = torch.stack(
+ [single_negative_image_embeds] * num_images_per_prompt, dim=0
+ )
+
+ if do_classifier_free_guidance:
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ single_image_embeds = single_image_embeds.to(device)
+
+ image_embeds.append(single_image_embeds)
+ else:
+ repeat_dims = [1]
+ image_embeds = []
+ for single_image_embeds in ip_adapter_image_embeds:
+ if do_classifier_free_guidance:
+ single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ single_negative_image_embeds = single_negative_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
+ )
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ else:
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ image_embeds.append(single_image_embeds)
+
+ return image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ image,
+ mask,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ip_adapter_image=None,
+ ip_adapter_image_embeds=None,
+ brushnet_conditioning_scale=1.0,
+ control_guidance_start=0.0,
+ control_guidance_end=1.0,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ # Check `image`
+ is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance(
+ self.brushnet, torch._dynamo.eval_frame.OptimizedModule
+ )
+ if (
+ isinstance(self.brushnet, BrushNetModel)
+ or is_compiled
+ and isinstance(self.brushnet._orig_mod, BrushNetModel)
+ ):
+ self.check_image(image, mask, prompt, prompt_embeds)
+ else:
+ assert False
+
+ # Check `brushnet_conditioning_scale`
+ if (
+ isinstance(self.brushnet, BrushNetModel)
+ or is_compiled
+ and isinstance(self.brushnet._orig_mod, BrushNetModel)
+ ):
+ if not isinstance(brushnet_conditioning_scale, float):
+ raise TypeError("For single brushnet: `brushnet_conditioning_scale` must be type `float`.")
+ else:
+ assert False
+
+ if not isinstance(control_guidance_start, (tuple, list)):
+ control_guidance_start = [control_guidance_start]
+
+ if not isinstance(control_guidance_end, (tuple, list)):
+ control_guidance_end = [control_guidance_end]
+
+ if len(control_guidance_start) != len(control_guidance_end):
+ raise ValueError(
+ f"`control_guidance_start` has {len(control_guidance_start)} elements, but `control_guidance_end` has {len(control_guidance_end)} elements. Make sure to provide the same number of elements to each list."
+ )
+
+ for start, end in zip(control_guidance_start, control_guidance_end):
+ if start >= end:
+ raise ValueError(
+ f"control guidance start: {start} cannot be larger or equal to control guidance end: {end}."
+ )
+ if start < 0.0:
+ raise ValueError(f"control guidance start: {start} can't be smaller than 0.")
+ if end > 1.0:
+ raise ValueError(f"control guidance end: {end} can't be larger than 1.0.")
+
+ if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
+ raise ValueError(
+ "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
+ )
+
+ if ip_adapter_image_embeds is not None:
+ if not isinstance(ip_adapter_image_embeds, list):
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
+ )
+ elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
+ )
+
+ def check_image(self, image, mask, prompt, prompt_embeds):
+ image_is_pil = isinstance(image, PIL.Image.Image)
+ image_is_tensor = isinstance(image, torch.Tensor)
+ image_is_np = isinstance(image, np.ndarray)
+ image_is_pil_list = isinstance(image, list) and isinstance(image[0], PIL.Image.Image)
+ image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor)
+ image_is_np_list = isinstance(image, list) and isinstance(image[0], np.ndarray)
+
+ if (
+ not image_is_pil
+ and not image_is_tensor
+ and not image_is_np
+ and not image_is_pil_list
+ and not image_is_tensor_list
+ and not image_is_np_list
+ ):
+ raise TypeError(
+ f"image must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(image)}"
+ )
+
+ mask_is_pil = isinstance(mask, PIL.Image.Image)
+ mask_is_tensor = isinstance(mask, torch.Tensor)
+ mask_is_np = isinstance(mask, np.ndarray)
+ mask_is_pil_list = isinstance(mask, list) and isinstance(mask[0], PIL.Image.Image)
+ mask_is_tensor_list = isinstance(mask, list) and isinstance(mask[0], torch.Tensor)
+ mask_is_np_list = isinstance(mask, list) and isinstance(mask[0], np.ndarray)
+
+ if (
+ not mask_is_pil
+ and not mask_is_tensor
+ and not mask_is_np
+ and not mask_is_pil_list
+ and not mask_is_tensor_list
+ and not mask_is_np_list
+ ):
+ raise TypeError(
+ f"mask must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(mask)}"
+ )
+
+ if image_is_pil:
+ image_batch_size = 1
+ else:
+ image_batch_size = len(image)
+
+ if prompt is not None and isinstance(prompt, str):
+ prompt_batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ prompt_batch_size = len(prompt)
+ elif prompt_embeds is not None:
+ prompt_batch_size = prompt_embeds.shape[0]
+
+ if image_batch_size != 1 and image_batch_size != prompt_batch_size:
+ raise ValueError(
+ f"If image batch size is not 1, image batch size must be same as prompt batch size. image batch size: {image_batch_size}, prompt batch size: {prompt_batch_size}"
+ )
+
+ def prepare_image(
+ self,
+ image,
+ width,
+ height,
+ batch_size,
+ num_images_per_prompt,
+ device,
+ dtype,
+ do_classifier_free_guidance=False,
+ guess_mode=False,
+ ):
+ image = self.image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32)
+ image_batch_size = image.shape[0]
+
+ if image_batch_size == 1:
+ repeat_by = batch_size
+ else:
+ # image batch size is the same as prompt batch size
+ repeat_by = num_images_per_prompt
+
+ image = image.repeat_interleave(repeat_by, dim=0)
+
+ image = image.to(device=device, dtype=dtype)
+
+ if do_classifier_free_guidance and not guess_mode:
+ image = torch.cat([image] * 2)
+
+ return image.to(device=device, dtype=dtype)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ noise = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = noise * self.scheduler.init_noise_sigma
+ return latents, noise
+
+ # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
+ def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
+ """
+ See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
+
+ Args:
+ timesteps (`torch.Tensor`):
+ generate embedding vectors at these timesteps
+ embedding_dim (`int`, *optional*, defaults to 512):
+ dimension of the embeddings to generate
+ dtype:
+ data type of the generated embeddings
+
+ Returns:
+ `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)`
+ """
+ assert len(w.shape) == 1
+ w = w * 1000.0
+
+ half_dim = embedding_dim // 2
+ emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
+ emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
+ emb = w.to(dtype)[:, None] * emb[None, :]
+ emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+ if embedding_dim % 2 == 1: # zero pad
+ emb = torch.nn.functional.pad(emb, (0, 1))
+ assert emb.shape == (w.shape[0], embedding_dim)
+ return emb
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def clip_skip(self):
+ return self._clip_skip
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
+
+ @property
+ def cross_attention_kwargs(self):
+ return self._cross_attention_kwargs
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ promptA: Union[str, List[str]] = None,
+ promptB: Union[str, List[str]] = None,
+ promptU: Union[str, List[str]] = None,
+ tradoff: float = 1.0,
+ tradoff_nag: float = 1.0,
+ image: PipelineImageInput = None,
+ mask: PipelineImageInput = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ timesteps: List[int] = None,
+ guidance_scale: float = 7.5,
+ negative_promptA: Optional[Union[str, List[str]]] = None,
+ negative_promptB: Optional[Union[str, List[str]]] = None,
+ negative_promptU: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ ip_adapter_image: Optional[PipelineImageInput] = None,
+ ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ brushnet_conditioning_scale: Union[float, List[float]] = 1.0,
+ guess_mode: bool = False,
+ control_guidance_start: Union[float, List[float]] = 0.0,
+ control_guidance_end: Union[float, List[float]] = 1.0,
+ clip_skip: Optional[int] = None,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ **kwargs,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, `List[np.ndarray]`,:
+ `List[List[torch.FloatTensor]]`, `List[List[np.ndarray]]` or `List[List[PIL.Image.Image]]`):
+ The BrushNet input condition to provide guidance to the `unet` for generation. If the type is
+ specified as `torch.FloatTensor`, it is passed to BrushNet as is. `PIL.Image.Image` can also be
+ accepted as an image. The dimensions of the output image defaults to `image`'s dimensions. If height
+ and/or width are passed, `image` is resized accordingly. If multiple BrushNets are specified in
+ `init`, images must be passed as a list such that each element of the list can be correctly batched for
+ input to a single BrushNet. When `prompt` is a list, and if a list of images is passed for a single BrushNet,
+ each will be paired with each prompt in the `prompt` list. This also applies to multiple BrushNets,
+ where a list of image lists can be passed to batch for each prompt and each BrushNet.
+ mask (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, `List[np.ndarray]`,:
+ `List[List[torch.FloatTensor]]`, `List[List[np.ndarray]]` or `List[List[PIL.Image.Image]]`):
+ The BrushNet input condition to provide guidance to the `unet` for generation. If the type is
+ specified as `torch.FloatTensor`, it is passed to BrushNet as is. `PIL.Image.Image` can also be
+ accepted as an image. The dimensions of the output image defaults to `image`'s dimensions. If height
+ and/or width are passed, `image` is resized accordingly. If multiple BrushNets are specified in
+ `init`, images must be passed as a list such that each element of the list can be correctly batched for
+ input to a single BrushNet. When `prompt` is a list, and if a list of images is passed for a single BrushNet,
+ each will be paired with each prompt in the `prompt` list. This also applies to multiple BrushNets,
+ where a list of image lists can be passed to batch for each prompt and each BrushNet.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+ in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+ passed will be used. Must be in descending order.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
+ ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
+ Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters.
+ Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding
+ if `do_classifier_free_guidance` is set to `True`.
+ If not provided, embeddings are computed from the `ip_adapter_image` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ brushnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0):
+ The outputs of the BrushNet are multiplied by `brushnet_conditioning_scale` before they are added
+ to the residual in the original `unet`. If multiple BrushNets are specified in `init`, you can set
+ the corresponding scale as a list.
+ guess_mode (`bool`, *optional*, defaults to `False`):
+ The BrushNet encoder tries to recognize the content of the input image even if you remove all
+ prompts. A `guidance_scale` value between 3.0 and 5.0 is recommended.
+ control_guidance_start (`float` or `List[float]`, *optional*, defaults to 0.0):
+ The percentage of total steps at which the BrushNet starts applying.
+ control_guidance_end (`float` or `List[float]`, *optional*, defaults to 1.0):
+ The percentage of total steps at which the BrushNet stops applying.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeine class.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images and the
+ second element is a list of `bool`s indicating whether the corresponding generated image contains
+ "not-safe-for-work" (nsfw) content.
+ """
+
+ callback = kwargs.pop("callback", None)
+ callback_steps = kwargs.pop("callback_steps", None)
+
+ if callback is not None:
+ deprecate(
+ "callback",
+ "1.0.0",
+ "Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+ )
+ if callback_steps is not None:
+ deprecate(
+ "callback_steps",
+ "1.0.0",
+ "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+ )
+
+ brushnet = self.brushnet._orig_mod if is_compiled_module(self.brushnet) else self.brushnet
+
+ # align format for control guidance
+ if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list):
+ control_guidance_start = len(control_guidance_end) * [control_guidance_start]
+ elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list):
+ control_guidance_end = len(control_guidance_start) * [control_guidance_end]
+ elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list):
+ control_guidance_start, control_guidance_end = (
+ [control_guidance_start],
+ [control_guidance_end],
+ )
+
+ # 1. Check inputs. Raise error if not correct
+ prompt = promptA
+ negative_prompt = negative_promptA
+ self.check_inputs(
+ prompt,
+ image,
+ mask,
+ callback_steps,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ brushnet_conditioning_scale,
+ control_guidance_start,
+ control_guidance_end,
+ callback_on_step_end_tensor_inputs,
+ )
+
+ self._guidance_scale = guidance_scale
+ self._clip_skip = clip_skip
+ self._cross_attention_kwargs = cross_attention_kwargs
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+
+ global_pool_conditions = (
+ brushnet.config.global_pool_conditions
+ if isinstance(brushnet, BrushNetModel)
+ else brushnet.nets[0].config.global_pool_conditions
+ )
+ guess_mode = guess_mode or global_pool_conditions
+
+ # 3. Encode input prompt
+ text_encoder_lora_scale = (
+ self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
+ )
+
+ prompt_embeds = self._encode_prompt(
+ promptA,
+ promptB,
+ tradoff,
+ device,
+ num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ negative_promptA,
+ negative_promptB,
+ tradoff_nag,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ )
+ prompt_embedsU = None
+ negative_prompt_embedsU = None
+ prompt_embedsU = self.encode_prompt(
+ promptU,
+ device,
+ num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ negative_promptU,
+ prompt_embeds=prompt_embedsU,
+ negative_prompt_embeds=negative_prompt_embedsU,
+ lora_scale=text_encoder_lora_scale,
+ )
+
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+ image_embeds = self.prepare_ip_adapter_image_embeds(
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ device,
+ batch_size * num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ )
+
+ # 4. Prepare image
+ if isinstance(brushnet, BrushNetModel):
+ image = self.prepare_image(
+ image=image,
+ width=width,
+ height=height,
+ batch_size=batch_size * num_images_per_prompt,
+ num_images_per_prompt=num_images_per_prompt,
+ device=device,
+ dtype=brushnet.dtype,
+ do_classifier_free_guidance=self.do_classifier_free_guidance,
+ guess_mode=guess_mode,
+ )
+ original_mask = self.prepare_image(
+ image=mask,
+ width=width,
+ height=height,
+ batch_size=batch_size * num_images_per_prompt,
+ num_images_per_prompt=num_images_per_prompt,
+ device=device,
+ dtype=brushnet.dtype,
+ do_classifier_free_guidance=self.do_classifier_free_guidance,
+ guess_mode=guess_mode,
+ )
+ original_mask = (original_mask.sum(1)[:, None, :, :] < 0).to(image.dtype)
+ height, width = image.shape[-2:]
+ else:
+ assert False
+
+ # 5. Prepare timesteps
+ timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+ self._num_timesteps = len(timesteps)
+
+ # 6. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ latents, noise = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 6.1 prepare condition latents
+ # mask_i = transforms.ToPILImage()(image[0:1,:,:,:].squeeze(0))
+ # mask_i.save('_mask.png')
+ # print(brushnet.dtype)
+ conditioning_latents = self.vae.encode(
+ image.to(device=device, dtype=brushnet.dtype)).latent_dist.sample() * self.vae.config.scaling_factor
+ mask = torch.nn.functional.interpolate(
+ original_mask,
+ size=(
+ conditioning_latents.shape[-2],
+ conditioning_latents.shape[-1]
+ )
+ )
+ conditioning_latents = torch.concat([conditioning_latents, mask], 1)
+ # image = self.vae.decode(conditioning_latents[:1,:4,:,:] / self.vae.config.scaling_factor, return_dict=False, generator=generator)[0]
+ # from torchvision import transforms
+ # mask_i = transforms.ToPILImage()(image[0:1,:,:,:].squeeze(0)/2+0.5)
+ # mask_i.save(str(timesteps[0]) +'_C.png')
+
+ # 6.5 Optionally get Guidance Scale Embedding
+ timestep_cond = None
+ if self.unet.config.time_cond_proj_dim is not None:
+ guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
+ timestep_cond = self.get_guidance_scale_embedding(
+ guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+ ).to(device=device, dtype=latents.dtype)
+
+ # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 7.1 Add image embeds for IP-Adapter
+ added_cond_kwargs = (
+ {"image_embeds": image_embeds}
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None
+ else None
+ )
+
+ # 7.2 Create tensor stating which brushnets to keep
+ brushnet_keep = []
+ for i in range(len(timesteps)):
+ keeps = [
+ 1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e)
+ for s, e in zip(control_guidance_start, control_guidance_end)
+ ]
+ brushnet_keep.append(keeps[0] if isinstance(brushnet, BrushNetModel) else keeps)
+
+ # 8. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ is_unet_compiled = is_compiled_module(self.unet)
+ is_brushnet_compiled = is_compiled_module(self.brushnet)
+ is_torch_higher_equal_2_1 = is_torch_version(">=", "2.1")
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # Relevant thread:
+ # https://dev-discuss.pytorch.org/t/cudagraphs-in-pytorch-2-0/1428
+ if (is_unet_compiled and is_brushnet_compiled) and is_torch_higher_equal_2_1:
+ torch._inductor.cudagraph_mark_step_begin()
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # brushnet(s) inference
+ if guess_mode and self.do_classifier_free_guidance:
+ # Infer BrushNet only for the conditional batch.
+ control_model_input = latents
+ control_model_input = self.scheduler.scale_model_input(control_model_input, t)
+ brushnet_prompt_embeds = prompt_embeds.chunk(2)[1]
+ else:
+ control_model_input = latent_model_input
+ brushnet_prompt_embeds = prompt_embeds
+
+ if isinstance(brushnet_keep[i], list):
+ cond_scale = [c * s for c, s in zip(brushnet_conditioning_scale, brushnet_keep[i])]
+ else:
+ brushnet_cond_scale = brushnet_conditioning_scale
+ if isinstance(brushnet_cond_scale, list):
+ brushnet_cond_scale = brushnet_cond_scale[0]
+ cond_scale = brushnet_cond_scale * brushnet_keep[i]
+
+ down_block_res_samples, mid_block_res_sample, up_block_res_samples = self.brushnet(
+ control_model_input,
+ t,
+ encoder_hidden_states=brushnet_prompt_embeds,
+ brushnet_cond=conditioning_latents,
+ conditioning_scale=cond_scale,
+ guess_mode=guess_mode,
+ return_dict=False,
+ )
+
+ if guess_mode and self.do_classifier_free_guidance:
+ # Infered BrushNet only for the conditional batch.
+ # To apply the output of BrushNet to both the unconditional and conditional batches,
+ # add 0 to the unconditional batch to keep it unchanged.
+ down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples]
+ mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample])
+ up_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in up_block_res_samples]
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embedsU,
+ timestep_cond=timestep_cond,
+ cross_attention_kwargs=self.cross_attention_kwargs,
+ down_block_add_samples=down_block_res_samples,
+ mid_block_add_sample=mid_block_res_sample,
+ up_block_add_samples=up_block_res_samples,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if self.do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+ negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ # If we do sequential model offloading, let's offload unet and brushnet
+ # manually for max memory savings
+ if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+ self.unet.to("cpu")
+ self.brushnet.to("cpu")
+ torch.cuda.empty_cache()
+
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[
+ 0
+ ]
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+ else:
+ image = latents
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/iopaint/model/power_paint/v2/unet_2d_blocks.py b/iopaint/model/power_paint/v2/unet_2d_blocks.py
new file mode 100644
index 0000000..228d0fc
--- /dev/null
+++ b/iopaint/model/power_paint/v2/unet_2d_blocks.py
@@ -0,0 +1,3811 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Any, Dict, Optional, Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from diffusers.utils import is_torch_version, logging
+from diffusers.utils.torch_utils import apply_freeu
+from diffusers.models.activations import get_activation
+from diffusers.models.attention_processor import Attention, AttnAddedKVProcessor, AttnAddedKVProcessor2_0
+from diffusers.models.normalization import AdaGroupNorm
+from diffusers.models.resnet import (
+ Downsample2D,
+ FirDownsample2D,
+ FirUpsample2D,
+ KDownsample2D,
+ KUpsample2D,
+ ResnetBlock2D,
+ ResnetBlockCondNorm2D,
+ Upsample2D,
+)
+from diffusers.models.transformers.dual_transformer_2d import DualTransformer2DModel
+from diffusers.models.transformers.transformer_2d import Transformer2DModel
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+def get_down_block(
+ down_block_type: str,
+ num_layers: int,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ add_downsample: bool,
+ resnet_eps: float,
+ resnet_act_fn: str,
+ transformer_layers_per_block: int = 1,
+ num_attention_heads: Optional[int] = None,
+ resnet_groups: Optional[int] = None,
+ cross_attention_dim: Optional[int] = None,
+ downsample_padding: Optional[int] = None,
+ dual_cross_attention: bool = False,
+ use_linear_projection: bool = False,
+ only_cross_attention: bool = False,
+ upcast_attention: bool = False,
+ resnet_time_scale_shift: str = "default",
+ attention_type: str = "default",
+ resnet_skip_time_act: bool = False,
+ resnet_out_scale_factor: float = 1.0,
+ cross_attention_norm: Optional[str] = None,
+ attention_head_dim: Optional[int] = None,
+ downsample_type: Optional[str] = None,
+ dropout: float = 0.0,
+):
+ # If attn head dim is not defined, we default it to the number of heads
+ if attention_head_dim is None:
+ logger.warn(
+ f"It is recommended to provide `attention_head_dim` when calling `get_down_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
+ )
+ attention_head_dim = num_attention_heads
+
+ down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type
+ if down_block_type == "DownBlock2D":
+ return DownBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ add_downsample=add_downsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ downsample_padding=downsample_padding,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ )
+ elif down_block_type == "ResnetDownsampleBlock2D":
+ return ResnetDownsampleBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ add_downsample=add_downsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ skip_time_act=resnet_skip_time_act,
+ output_scale_factor=resnet_out_scale_factor,
+ )
+ elif down_block_type == "AttnDownBlock2D":
+ if add_downsample is False:
+ downsample_type = None
+ else:
+ downsample_type = downsample_type or "conv" # default to 'conv'
+ return AttnDownBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ downsample_padding=downsample_padding,
+ attention_head_dim=attention_head_dim,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ downsample_type=downsample_type,
+ )
+ elif down_block_type == "CrossAttnDownBlock2D":
+ if cross_attention_dim is None:
+ raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlock2D")
+ return CrossAttnDownBlock2D(
+ num_layers=num_layers,
+ transformer_layers_per_block=transformer_layers_per_block,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ add_downsample=add_downsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ downsample_padding=downsample_padding,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=num_attention_heads,
+ dual_cross_attention=dual_cross_attention,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention,
+ upcast_attention=upcast_attention,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ attention_type=attention_type,
+ )
+ elif down_block_type == "SimpleCrossAttnDownBlock2D":
+ if cross_attention_dim is None:
+ raise ValueError("cross_attention_dim must be specified for SimpleCrossAttnDownBlock2D")
+ return SimpleCrossAttnDownBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ add_downsample=add_downsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ cross_attention_dim=cross_attention_dim,
+ attention_head_dim=attention_head_dim,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ skip_time_act=resnet_skip_time_act,
+ output_scale_factor=resnet_out_scale_factor,
+ only_cross_attention=only_cross_attention,
+ cross_attention_norm=cross_attention_norm,
+ )
+ elif down_block_type == "SkipDownBlock2D":
+ return SkipDownBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ add_downsample=add_downsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ downsample_padding=downsample_padding,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ )
+ elif down_block_type == "AttnSkipDownBlock2D":
+ return AttnSkipDownBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ add_downsample=add_downsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ attention_head_dim=attention_head_dim,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ )
+ elif down_block_type == "DownEncoderBlock2D":
+ return DownEncoderBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ dropout=dropout,
+ add_downsample=add_downsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ downsample_padding=downsample_padding,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ )
+ elif down_block_type == "AttnDownEncoderBlock2D":
+ return AttnDownEncoderBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ dropout=dropout,
+ add_downsample=add_downsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ downsample_padding=downsample_padding,
+ attention_head_dim=attention_head_dim,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ )
+ elif down_block_type == "KDownBlock2D":
+ return KDownBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ add_downsample=add_downsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ )
+ elif down_block_type == "KCrossAttnDownBlock2D":
+ return KCrossAttnDownBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ add_downsample=add_downsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ cross_attention_dim=cross_attention_dim,
+ attention_head_dim=attention_head_dim,
+ add_self_attention=True if not add_downsample else False,
+ )
+ raise ValueError(f"{down_block_type} does not exist.")
+
+
+def get_mid_block(
+ mid_block_type: str,
+ temb_channels: int,
+ in_channels: int,
+ resnet_eps: float,
+ resnet_act_fn: str,
+ resnet_groups: int,
+ output_scale_factor: float = 1.0,
+ transformer_layers_per_block: int = 1,
+ num_attention_heads: Optional[int] = None,
+ cross_attention_dim: Optional[int] = None,
+ dual_cross_attention: bool = False,
+ use_linear_projection: bool = False,
+ mid_block_only_cross_attention: bool = False,
+ upcast_attention: bool = False,
+ resnet_time_scale_shift: str = "default",
+ attention_type: str = "default",
+ resnet_skip_time_act: bool = False,
+ cross_attention_norm: Optional[str] = None,
+ attention_head_dim: Optional[int] = 1,
+ dropout: float = 0.0,
+):
+ if mid_block_type == "UNetMidBlock2DCrossAttn":
+ return UNetMidBlock2DCrossAttn(
+ transformer_layers_per_block=transformer_layers_per_block,
+ in_channels=in_channels,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=num_attention_heads,
+ resnet_groups=resnet_groups,
+ dual_cross_attention=dual_cross_attention,
+ use_linear_projection=use_linear_projection,
+ upcast_attention=upcast_attention,
+ attention_type=attention_type,
+ )
+ elif mid_block_type == "UNetMidBlock2DSimpleCrossAttn":
+ return UNetMidBlock2DSimpleCrossAttn(
+ in_channels=in_channels,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ cross_attention_dim=cross_attention_dim,
+ attention_head_dim=attention_head_dim,
+ resnet_groups=resnet_groups,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ skip_time_act=resnet_skip_time_act,
+ only_cross_attention=mid_block_only_cross_attention,
+ cross_attention_norm=cross_attention_norm,
+ )
+ elif mid_block_type == "UNetMidBlock2D":
+ return UNetMidBlock2D(
+ in_channels=in_channels,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ num_layers=0,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ resnet_groups=resnet_groups,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ add_attention=False,
+ )
+ elif mid_block_type == "MidBlock2D":
+ return MidBlock2D(
+ in_channels=in_channels,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ resnet_groups=resnet_groups,
+ use_linear_projection=use_linear_projection,
+ )
+ elif mid_block_type is None:
+ return None
+ else:
+ raise ValueError(f"unknown mid_block_type : {mid_block_type}")
+
+
+def get_up_block(
+ up_block_type: str,
+ num_layers: int,
+ in_channels: int,
+ out_channels: int,
+ prev_output_channel: int,
+ temb_channels: int,
+ add_upsample: bool,
+ resnet_eps: float,
+ resnet_act_fn: str,
+ resolution_idx: Optional[int] = None,
+ transformer_layers_per_block: int = 1,
+ num_attention_heads: Optional[int] = None,
+ resnet_groups: Optional[int] = None,
+ cross_attention_dim: Optional[int] = None,
+ dual_cross_attention: bool = False,
+ use_linear_projection: bool = False,
+ only_cross_attention: bool = False,
+ upcast_attention: bool = False,
+ resnet_time_scale_shift: str = "default",
+ attention_type: str = "default",
+ resnet_skip_time_act: bool = False,
+ resnet_out_scale_factor: float = 1.0,
+ cross_attention_norm: Optional[str] = None,
+ attention_head_dim: Optional[int] = None,
+ upsample_type: Optional[str] = None,
+ dropout: float = 0.0,
+) -> nn.Module:
+ # If attn head dim is not defined, we default it to the number of heads
+ if attention_head_dim is None:
+ logger.warn(
+ f"It is recommended to provide `attention_head_dim` when calling `get_up_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
+ )
+ attention_head_dim = num_attention_heads
+
+ up_block_type = up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
+ if up_block_type == "UpBlock2D":
+ return UpBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ prev_output_channel=prev_output_channel,
+ temb_channels=temb_channels,
+ resolution_idx=resolution_idx,
+ dropout=dropout,
+ add_upsample=add_upsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ )
+ elif up_block_type == "ResnetUpsampleBlock2D":
+ return ResnetUpsampleBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ prev_output_channel=prev_output_channel,
+ temb_channels=temb_channels,
+ resolution_idx=resolution_idx,
+ dropout=dropout,
+ add_upsample=add_upsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ skip_time_act=resnet_skip_time_act,
+ output_scale_factor=resnet_out_scale_factor,
+ )
+ elif up_block_type == "CrossAttnUpBlock2D":
+ if cross_attention_dim is None:
+ raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlock2D")
+ return CrossAttnUpBlock2D(
+ num_layers=num_layers,
+ transformer_layers_per_block=transformer_layers_per_block,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ prev_output_channel=prev_output_channel,
+ temb_channels=temb_channels,
+ resolution_idx=resolution_idx,
+ dropout=dropout,
+ add_upsample=add_upsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=num_attention_heads,
+ dual_cross_attention=dual_cross_attention,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention,
+ upcast_attention=upcast_attention,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ attention_type=attention_type,
+ )
+ elif up_block_type == "SimpleCrossAttnUpBlock2D":
+ if cross_attention_dim is None:
+ raise ValueError("cross_attention_dim must be specified for SimpleCrossAttnUpBlock2D")
+ return SimpleCrossAttnUpBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ prev_output_channel=prev_output_channel,
+ temb_channels=temb_channels,
+ resolution_idx=resolution_idx,
+ dropout=dropout,
+ add_upsample=add_upsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ cross_attention_dim=cross_attention_dim,
+ attention_head_dim=attention_head_dim,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ skip_time_act=resnet_skip_time_act,
+ output_scale_factor=resnet_out_scale_factor,
+ only_cross_attention=only_cross_attention,
+ cross_attention_norm=cross_attention_norm,
+ )
+ elif up_block_type == "AttnUpBlock2D":
+ if add_upsample is False:
+ upsample_type = None
+ else:
+ upsample_type = upsample_type or "conv" # default to 'conv'
+
+ return AttnUpBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ prev_output_channel=prev_output_channel,
+ temb_channels=temb_channels,
+ resolution_idx=resolution_idx,
+ dropout=dropout,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ attention_head_dim=attention_head_dim,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ upsample_type=upsample_type,
+ )
+ elif up_block_type == "SkipUpBlock2D":
+ return SkipUpBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ prev_output_channel=prev_output_channel,
+ temb_channels=temb_channels,
+ resolution_idx=resolution_idx,
+ dropout=dropout,
+ add_upsample=add_upsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ )
+ elif up_block_type == "AttnSkipUpBlock2D":
+ return AttnSkipUpBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ prev_output_channel=prev_output_channel,
+ temb_channels=temb_channels,
+ resolution_idx=resolution_idx,
+ dropout=dropout,
+ add_upsample=add_upsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ attention_head_dim=attention_head_dim,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ )
+ elif up_block_type == "UpDecoderBlock2D":
+ return UpDecoderBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ resolution_idx=resolution_idx,
+ dropout=dropout,
+ add_upsample=add_upsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ temb_channels=temb_channels,
+ )
+ elif up_block_type == "AttnUpDecoderBlock2D":
+ return AttnUpDecoderBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ resolution_idx=resolution_idx,
+ dropout=dropout,
+ add_upsample=add_upsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ attention_head_dim=attention_head_dim,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ temb_channels=temb_channels,
+ )
+ elif up_block_type == "KUpBlock2D":
+ return KUpBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ resolution_idx=resolution_idx,
+ dropout=dropout,
+ add_upsample=add_upsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ )
+ elif up_block_type == "KCrossAttnUpBlock2D":
+ return KCrossAttnUpBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ resolution_idx=resolution_idx,
+ dropout=dropout,
+ add_upsample=add_upsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ cross_attention_dim=cross_attention_dim,
+ attention_head_dim=attention_head_dim,
+ )
+
+ raise ValueError(f"{up_block_type} does not exist.")
+
+
+class AutoencoderTinyBlock(nn.Module):
+ """
+ Tiny Autoencoder block used in [`AutoencoderTiny`]. It is a mini residual module consisting of plain conv + ReLU
+ blocks.
+
+ Args:
+ in_channels (`int`): The number of input channels.
+ out_channels (`int`): The number of output channels.
+ act_fn (`str`):
+ ` The activation function to use. Supported values are `"swish"`, `"mish"`, `"gelu"`, and `"relu"`.
+
+ Returns:
+ `torch.FloatTensor`: A tensor with the same shape as the input tensor, but with the number of channels equal to
+ `out_channels`.
+ """
+
+ def __init__(self, in_channels: int, out_channels: int, act_fn: str):
+ super().__init__()
+ act_fn = get_activation(act_fn)
+ self.conv = nn.Sequential(
+ nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1),
+ act_fn,
+ nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),
+ act_fn,
+ nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),
+ )
+ self.skip = (
+ nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False)
+ if in_channels != out_channels
+ else nn.Identity()
+ )
+ self.fuse = nn.ReLU()
+
+ def forward(self, x: torch.FloatTensor) -> torch.FloatTensor:
+ return self.fuse(self.conv(x) + self.skip(x))
+
+
+class UNetMidBlock2D(nn.Module):
+ """
+ A 2D UNet mid-block [`UNetMidBlock2D`] with multiple residual blocks and optional attention blocks.
+
+ Args:
+ in_channels (`int`): The number of input channels.
+ temb_channels (`int`): The number of temporal embedding channels.
+ dropout (`float`, *optional*, defaults to 0.0): The dropout rate.
+ num_layers (`int`, *optional*, defaults to 1): The number of residual blocks.
+ resnet_eps (`float`, *optional*, 1e-6 ): The epsilon value for the resnet blocks.
+ resnet_time_scale_shift (`str`, *optional*, defaults to `default`):
+ The type of normalization to apply to the time embeddings. This can help to improve the performance of the
+ model on tasks with long-range temporal dependencies.
+ resnet_act_fn (`str`, *optional*, defaults to `swish`): The activation function for the resnet blocks.
+ resnet_groups (`int`, *optional*, defaults to 32):
+ The number of groups to use in the group normalization layers of the resnet blocks.
+ attn_groups (`Optional[int]`, *optional*, defaults to None): The number of groups for the attention blocks.
+ resnet_pre_norm (`bool`, *optional*, defaults to `True`):
+ Whether to use pre-normalization for the resnet blocks.
+ add_attention (`bool`, *optional*, defaults to `True`): Whether to add attention blocks.
+ attention_head_dim (`int`, *optional*, defaults to 1):
+ Dimension of a single attention head. The number of attention heads is determined based on this value and
+ the number of input channels.
+ output_scale_factor (`float`, *optional*, defaults to 1.0): The output scale factor.
+
+ Returns:
+ `torch.FloatTensor`: The output of the last residual block, which is a tensor of shape `(batch_size,
+ in_channels, height, width)`.
+
+ """
+
+ def __init__(
+ self,
+ in_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default", # default, spatial
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ attn_groups: Optional[int] = None,
+ resnet_pre_norm: bool = True,
+ add_attention: bool = True,
+ attention_head_dim: int = 1,
+ output_scale_factor: float = 1.0,
+ ):
+ super().__init__()
+ resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+ self.add_attention = add_attention
+
+ if attn_groups is None:
+ attn_groups = resnet_groups if resnet_time_scale_shift == "default" else None
+
+ # there is always at least one resnet
+ if resnet_time_scale_shift == "spatial":
+ resnets = [
+ ResnetBlockCondNorm2D(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm="spatial",
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ )
+ ]
+ else:
+ resnets = [
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ ]
+ attentions = []
+
+ if attention_head_dim is None:
+ logger.warn(
+ f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {in_channels}."
+ )
+ attention_head_dim = in_channels
+
+ for _ in range(num_layers):
+ if self.add_attention:
+ attentions.append(
+ Attention(
+ in_channels,
+ heads=in_channels // attention_head_dim,
+ dim_head=attention_head_dim,
+ rescale_output_factor=output_scale_factor,
+ eps=resnet_eps,
+ norm_num_groups=attn_groups,
+ spatial_norm_dim=temb_channels if resnet_time_scale_shift == "spatial" else None,
+ residual_connection=True,
+ bias=True,
+ upcast_softmax=True,
+ _from_deprecated_attn_block=True,
+ )
+ )
+ else:
+ attentions.append(None)
+
+ if resnet_time_scale_shift == "spatial":
+ resnets.append(
+ ResnetBlockCondNorm2D(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm="spatial",
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ )
+ )
+ else:
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ def forward(self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None) -> torch.FloatTensor:
+ hidden_states = self.resnets[0](hidden_states, temb)
+ for attn, resnet in zip(self.attentions, self.resnets[1:]):
+ if attn is not None:
+ hidden_states = attn(hidden_states, temb=temb)
+ hidden_states = resnet(hidden_states, temb)
+
+ return hidden_states
+
+
+class UNetMidBlock2DCrossAttn(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ num_attention_heads: int = 1,
+ output_scale_factor: float = 1.0,
+ cross_attention_dim: int = 1280,
+ dual_cross_attention: bool = False,
+ use_linear_projection: bool = False,
+ upcast_attention: bool = False,
+ attention_type: str = "default",
+ ):
+ super().__init__()
+
+ self.has_cross_attention = True
+ self.num_attention_heads = num_attention_heads
+ resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+
+ # support for variable transformer layers per block
+ if isinstance(transformer_layers_per_block, int):
+ transformer_layers_per_block = [transformer_layers_per_block] * num_layers
+
+ # there is always at least one resnet
+ resnets = [
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ ]
+ attentions = []
+
+ for i in range(num_layers):
+ if not dual_cross_attention:
+ attentions.append(
+ Transformer2DModel(
+ num_attention_heads,
+ in_channels // num_attention_heads,
+ in_channels=in_channels,
+ num_layers=transformer_layers_per_block[i],
+ cross_attention_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ use_linear_projection=use_linear_projection,
+ upcast_attention=upcast_attention,
+ attention_type=attention_type,
+ )
+ )
+ else:
+ attentions.append(
+ DualTransformer2DModel(
+ num_attention_heads,
+ in_channels // num_attention_heads,
+ in_channels=in_channels,
+ num_layers=1,
+ cross_attention_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ )
+ )
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+ hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale)
+ for attn, resnet in zip(self.attentions, self.resnets[1:]):
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module, return_dict=None):
+ def custom_forward(*inputs):
+ if return_dict is not None:
+ return module(*inputs, return_dict=return_dict)
+ else:
+ return module(*inputs)
+
+ return custom_forward
+
+ ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ attention_mask=attention_mask,
+ encoder_attention_mask=encoder_attention_mask,
+ return_dict=False,
+ )[0]
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet),
+ hidden_states,
+ temb,
+ **ckpt_kwargs,
+ )
+ else:
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ attention_mask=attention_mask,
+ encoder_attention_mask=encoder_attention_mask,
+ return_dict=False,
+ )[0]
+ hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+
+ return hidden_states
+
+
+class UNetMidBlock2DSimpleCrossAttn(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ attention_head_dim: int = 1,
+ output_scale_factor: float = 1.0,
+ cross_attention_dim: int = 1280,
+ skip_time_act: bool = False,
+ only_cross_attention: bool = False,
+ cross_attention_norm: Optional[str] = None,
+ ):
+ super().__init__()
+
+ self.has_cross_attention = True
+
+ self.attention_head_dim = attention_head_dim
+ resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+
+ self.num_heads = in_channels // self.attention_head_dim
+
+ # there is always at least one resnet
+ resnets = [
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ skip_time_act=skip_time_act,
+ )
+ ]
+ attentions = []
+
+ for _ in range(num_layers):
+ processor = (
+ AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor()
+ )
+
+ attentions.append(
+ Attention(
+ query_dim=in_channels,
+ cross_attention_dim=in_channels,
+ heads=self.num_heads,
+ dim_head=self.attention_head_dim,
+ added_kv_proj_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ bias=True,
+ upcast_softmax=True,
+ only_cross_attention=only_cross_attention,
+ cross_attention_norm=cross_attention_norm,
+ processor=processor,
+ )
+ )
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ skip_time_act=skip_time_act,
+ )
+ )
+
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
+ lora_scale = cross_attention_kwargs.get("scale", 1.0)
+
+ if attention_mask is None:
+ # if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask.
+ mask = None if encoder_hidden_states is None else encoder_attention_mask
+ else:
+ # when attention_mask is defined: we don't even check for encoder_attention_mask.
+ # this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks.
+ # TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask.
+ # then we can simplify this whole if/else block to:
+ # mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask
+ mask = attention_mask
+
+ hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale)
+ for attn, resnet in zip(self.attentions, self.resnets[1:]):
+ # attn
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=mask,
+ **cross_attention_kwargs,
+ )
+
+ # resnet
+ hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+
+ return hidden_states
+
+
+class MidBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ output_scale_factor: float = 1.0,
+ use_linear_projection: bool = False,
+ ):
+ super().__init__()
+
+ self.has_cross_attention = False
+ resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+
+ # there is always at least one resnet
+ resnets = [
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ ]
+
+ for i in range(num_layers):
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+
+ self.resnets = nn.ModuleList(resnets)
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ lora_scale = 1.0
+ hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale)
+ for resnet in self.resnets[1:]:
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module, return_dict=None):
+ def custom_forward(*inputs):
+ if return_dict is not None:
+ return module(*inputs, return_dict=return_dict)
+ else:
+ return module(*inputs)
+
+ return custom_forward
+
+ ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet),
+ hidden_states,
+ temb,
+ **ckpt_kwargs,
+ )
+ else:
+ hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+
+ return hidden_states
+
+
+class AttnDownBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ attention_head_dim: int = 1,
+ output_scale_factor: float = 1.0,
+ downsample_padding: int = 1,
+ downsample_type: str = "conv",
+ ):
+ super().__init__()
+ resnets = []
+ attentions = []
+ self.downsample_type = downsample_type
+
+ if attention_head_dim is None:
+ logger.warn(
+ f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}."
+ )
+ attention_head_dim = out_channels
+
+ for i in range(num_layers):
+ in_channels = in_channels if i == 0 else out_channels
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+ attentions.append(
+ Attention(
+ out_channels,
+ heads=out_channels // attention_head_dim,
+ dim_head=attention_head_dim,
+ rescale_output_factor=output_scale_factor,
+ eps=resnet_eps,
+ norm_num_groups=resnet_groups,
+ residual_connection=True,
+ bias=True,
+ upcast_softmax=True,
+ _from_deprecated_attn_block=True,
+ )
+ )
+
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ if downsample_type == "conv":
+ self.downsamplers = nn.ModuleList(
+ [
+ Downsample2D(
+ out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
+ )
+ ]
+ )
+ elif downsample_type == "resnet":
+ self.downsamplers = nn.ModuleList(
+ [
+ ResnetBlock2D(
+ in_channels=out_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ down=True,
+ )
+ ]
+ )
+ else:
+ self.downsamplers = None
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ upsample_size: Optional[int] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
+ cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
+
+ lora_scale = cross_attention_kwargs.get("scale", 1.0)
+
+ output_states = ()
+
+ for resnet, attn in zip(self.resnets, self.attentions):
+ cross_attention_kwargs.update({"scale": lora_scale})
+ hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+ hidden_states = attn(hidden_states, **cross_attention_kwargs)
+ output_states = output_states + (hidden_states,)
+
+ if self.downsamplers is not None:
+ for downsampler in self.downsamplers:
+ if self.downsample_type == "resnet":
+ hidden_states = downsampler(hidden_states, temb=temb, scale=lora_scale)
+ else:
+ hidden_states = downsampler(hidden_states, scale=lora_scale)
+
+ output_states += (hidden_states,)
+
+ return hidden_states, output_states
+
+
+class CrossAttnDownBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ num_attention_heads: int = 1,
+ cross_attention_dim: int = 1280,
+ output_scale_factor: float = 1.0,
+ downsample_padding: int = 1,
+ add_downsample: bool = True,
+ dual_cross_attention: bool = False,
+ use_linear_projection: bool = False,
+ only_cross_attention: bool = False,
+ upcast_attention: bool = False,
+ attention_type: str = "default",
+ ):
+ super().__init__()
+ resnets = []
+ attentions = []
+
+ self.has_cross_attention = True
+ self.num_attention_heads = num_attention_heads
+ if isinstance(transformer_layers_per_block, int):
+ transformer_layers_per_block = [transformer_layers_per_block] * num_layers
+
+ for i in range(num_layers):
+ in_channels = in_channels if i == 0 else out_channels
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+ if not dual_cross_attention:
+ attentions.append(
+ Transformer2DModel(
+ num_attention_heads,
+ out_channels // num_attention_heads,
+ in_channels=out_channels,
+ num_layers=transformer_layers_per_block[i],
+ cross_attention_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention,
+ upcast_attention=upcast_attention,
+ attention_type=attention_type,
+ )
+ )
+ else:
+ attentions.append(
+ DualTransformer2DModel(
+ num_attention_heads,
+ out_channels // num_attention_heads,
+ in_channels=out_channels,
+ num_layers=1,
+ cross_attention_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ )
+ )
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_downsample:
+ self.downsamplers = nn.ModuleList(
+ [
+ Downsample2D(
+ out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
+ )
+ ]
+ )
+ else:
+ self.downsamplers = None
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
+ additional_residuals: Optional[torch.FloatTensor] = None,
+ down_block_add_samples: Optional[torch.FloatTensor] = None,
+ ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
+ output_states = ()
+
+ lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+
+ blocks = list(zip(self.resnets, self.attentions))
+
+ for i, (resnet, attn) in enumerate(blocks):
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module, return_dict=None):
+ def custom_forward(*inputs):
+ if return_dict is not None:
+ return module(*inputs, return_dict=return_dict)
+ else:
+ return module(*inputs)
+
+ return custom_forward
+
+ ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet),
+ hidden_states,
+ temb,
+ **ckpt_kwargs,
+ )
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ attention_mask=attention_mask,
+ encoder_attention_mask=encoder_attention_mask,
+ return_dict=False,
+ )[0]
+ else:
+ hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ attention_mask=attention_mask,
+ encoder_attention_mask=encoder_attention_mask,
+ return_dict=False,
+ )[0]
+
+ # apply additional residuals to the output of the last pair of resnet and attention blocks
+ if i == len(blocks) - 1 and additional_residuals is not None:
+ hidden_states = hidden_states + additional_residuals
+
+ if down_block_add_samples is not None:
+ hidden_states = hidden_states + down_block_add_samples.pop(0)
+
+ output_states = output_states + (hidden_states,)
+
+ if self.downsamplers is not None:
+ for downsampler in self.downsamplers:
+ hidden_states = downsampler(hidden_states, scale=lora_scale)
+
+ if down_block_add_samples is not None:
+ hidden_states = hidden_states + down_block_add_samples.pop(0) # todo: add before or after
+
+ output_states = output_states + (hidden_states,)
+
+ return hidden_states, output_states
+
+
+class DownBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ output_scale_factor: float = 1.0,
+ add_downsample: bool = True,
+ downsample_padding: int = 1,
+ ):
+ super().__init__()
+ resnets = []
+
+ for i in range(num_layers):
+ in_channels = in_channels if i == 0 else out_channels
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_downsample:
+ self.downsamplers = nn.ModuleList(
+ [
+ Downsample2D(
+ out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
+ )
+ ]
+ )
+ else:
+ self.downsamplers = None
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0,
+ down_block_add_samples: Optional[torch.FloatTensor] = None,
+ ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
+ output_states = ()
+
+ for resnet in self.resnets:
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module):
+ def custom_forward(*inputs):
+ return module(*inputs)
+
+ return custom_forward
+
+ if is_torch_version(">=", "1.11.0"):
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
+ )
+ else:
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet), hidden_states, temb
+ )
+ else:
+ hidden_states = resnet(hidden_states, temb, scale=scale)
+
+ if down_block_add_samples is not None:
+ hidden_states = hidden_states + down_block_add_samples.pop(0)
+
+ output_states = output_states + (hidden_states,)
+
+ if self.downsamplers is not None:
+ for downsampler in self.downsamplers:
+ hidden_states = downsampler(hidden_states, scale=scale)
+
+ if down_block_add_samples is not None:
+ hidden_states = hidden_states + down_block_add_samples.pop(0) # todo: add before or after
+
+ output_states = output_states + (hidden_states,)
+
+ return hidden_states, output_states
+
+
+class DownEncoderBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ output_scale_factor: float = 1.0,
+ add_downsample: bool = True,
+ downsample_padding: int = 1,
+ ):
+ super().__init__()
+ resnets = []
+
+ for i in range(num_layers):
+ in_channels = in_channels if i == 0 else out_channels
+ if resnet_time_scale_shift == "spatial":
+ resnets.append(
+ ResnetBlockCondNorm2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=None,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm="spatial",
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ )
+ )
+ else:
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=None,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_downsample:
+ self.downsamplers = nn.ModuleList(
+ [
+ Downsample2D(
+ out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
+ )
+ ]
+ )
+ else:
+ self.downsamplers = None
+
+ def forward(self, hidden_states: torch.FloatTensor, scale: float = 1.0) -> torch.FloatTensor:
+ for resnet in self.resnets:
+ hidden_states = resnet(hidden_states, temb=None, scale=scale)
+
+ if self.downsamplers is not None:
+ for downsampler in self.downsamplers:
+ hidden_states = downsampler(hidden_states, scale)
+
+ return hidden_states
+
+
+class AttnDownEncoderBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ attention_head_dim: int = 1,
+ output_scale_factor: float = 1.0,
+ add_downsample: bool = True,
+ downsample_padding: int = 1,
+ ):
+ super().__init__()
+ resnets = []
+ attentions = []
+
+ if attention_head_dim is None:
+ logger.warn(
+ f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}."
+ )
+ attention_head_dim = out_channels
+
+ for i in range(num_layers):
+ in_channels = in_channels if i == 0 else out_channels
+ if resnet_time_scale_shift == "spatial":
+ resnets.append(
+ ResnetBlockCondNorm2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=None,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm="spatial",
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ )
+ )
+ else:
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=None,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+ attentions.append(
+ Attention(
+ out_channels,
+ heads=out_channels // attention_head_dim,
+ dim_head=attention_head_dim,
+ rescale_output_factor=output_scale_factor,
+ eps=resnet_eps,
+ norm_num_groups=resnet_groups,
+ residual_connection=True,
+ bias=True,
+ upcast_softmax=True,
+ _from_deprecated_attn_block=True,
+ )
+ )
+
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_downsample:
+ self.downsamplers = nn.ModuleList(
+ [
+ Downsample2D(
+ out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
+ )
+ ]
+ )
+ else:
+ self.downsamplers = None
+
+ def forward(self, hidden_states: torch.FloatTensor, scale: float = 1.0) -> torch.FloatTensor:
+ for resnet, attn in zip(self.resnets, self.attentions):
+ hidden_states = resnet(hidden_states, temb=None, scale=scale)
+ cross_attention_kwargs = {"scale": scale}
+ hidden_states = attn(hidden_states, **cross_attention_kwargs)
+
+ if self.downsamplers is not None:
+ for downsampler in self.downsamplers:
+ hidden_states = downsampler(hidden_states, scale)
+
+ return hidden_states
+
+
+class AttnSkipDownBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_pre_norm: bool = True,
+ attention_head_dim: int = 1,
+ output_scale_factor: float = np.sqrt(2.0),
+ add_downsample: bool = True,
+ ):
+ super().__init__()
+ self.attentions = nn.ModuleList([])
+ self.resnets = nn.ModuleList([])
+
+ if attention_head_dim is None:
+ logger.warn(
+ f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}."
+ )
+ attention_head_dim = out_channels
+
+ for i in range(num_layers):
+ in_channels = in_channels if i == 0 else out_channels
+ self.resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=min(in_channels // 4, 32),
+ groups_out=min(out_channels // 4, 32),
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+ self.attentions.append(
+ Attention(
+ out_channels,
+ heads=out_channels // attention_head_dim,
+ dim_head=attention_head_dim,
+ rescale_output_factor=output_scale_factor,
+ eps=resnet_eps,
+ norm_num_groups=32,
+ residual_connection=True,
+ bias=True,
+ upcast_softmax=True,
+ _from_deprecated_attn_block=True,
+ )
+ )
+
+ if add_downsample:
+ self.resnet_down = ResnetBlock2D(
+ in_channels=out_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=min(out_channels // 4, 32),
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ use_in_shortcut=True,
+ down=True,
+ kernel="fir",
+ )
+ self.downsamplers = nn.ModuleList([FirDownsample2D(out_channels, out_channels=out_channels)])
+ self.skip_conv = nn.Conv2d(3, out_channels, kernel_size=(1, 1), stride=(1, 1))
+ else:
+ self.resnet_down = None
+ self.downsamplers = None
+ self.skip_conv = None
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ skip_sample: Optional[torch.FloatTensor] = None,
+ scale: float = 1.0,
+ ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...], torch.FloatTensor]:
+ output_states = ()
+
+ for resnet, attn in zip(self.resnets, self.attentions):
+ hidden_states = resnet(hidden_states, temb, scale=scale)
+ cross_attention_kwargs = {"scale": scale}
+ hidden_states = attn(hidden_states, **cross_attention_kwargs)
+ output_states += (hidden_states,)
+
+ if self.downsamplers is not None:
+ hidden_states = self.resnet_down(hidden_states, temb, scale=scale)
+ for downsampler in self.downsamplers:
+ skip_sample = downsampler(skip_sample)
+
+ hidden_states = self.skip_conv(skip_sample) + hidden_states
+
+ output_states += (hidden_states,)
+
+ return hidden_states, output_states, skip_sample
+
+
+class SkipDownBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_pre_norm: bool = True,
+ output_scale_factor: float = np.sqrt(2.0),
+ add_downsample: bool = True,
+ downsample_padding: int = 1,
+ ):
+ super().__init__()
+ self.resnets = nn.ModuleList([])
+
+ for i in range(num_layers):
+ in_channels = in_channels if i == 0 else out_channels
+ self.resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=min(in_channels // 4, 32),
+ groups_out=min(out_channels // 4, 32),
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+
+ if add_downsample:
+ self.resnet_down = ResnetBlock2D(
+ in_channels=out_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=min(out_channels // 4, 32),
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ use_in_shortcut=True,
+ down=True,
+ kernel="fir",
+ )
+ self.downsamplers = nn.ModuleList([FirDownsample2D(out_channels, out_channels=out_channels)])
+ self.skip_conv = nn.Conv2d(3, out_channels, kernel_size=(1, 1), stride=(1, 1))
+ else:
+ self.resnet_down = None
+ self.downsamplers = None
+ self.skip_conv = None
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ skip_sample: Optional[torch.FloatTensor] = None,
+ scale: float = 1.0,
+ ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...], torch.FloatTensor]:
+ output_states = ()
+
+ for resnet in self.resnets:
+ hidden_states = resnet(hidden_states, temb, scale)
+ output_states += (hidden_states,)
+
+ if self.downsamplers is not None:
+ hidden_states = self.resnet_down(hidden_states, temb, scale)
+ for downsampler in self.downsamplers:
+ skip_sample = downsampler(skip_sample)
+
+ hidden_states = self.skip_conv(skip_sample) + hidden_states
+
+ output_states += (hidden_states,)
+
+ return hidden_states, output_states, skip_sample
+
+
+class ResnetDownsampleBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ output_scale_factor: float = 1.0,
+ add_downsample: bool = True,
+ skip_time_act: bool = False,
+ ):
+ super().__init__()
+ resnets = []
+
+ for i in range(num_layers):
+ in_channels = in_channels if i == 0 else out_channels
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ skip_time_act=skip_time_act,
+ )
+ )
+
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_downsample:
+ self.downsamplers = nn.ModuleList(
+ [
+ ResnetBlock2D(
+ in_channels=out_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ skip_time_act=skip_time_act,
+ down=True,
+ )
+ ]
+ )
+ else:
+ self.downsamplers = None
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0
+ ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
+ output_states = ()
+
+ for resnet in self.resnets:
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module):
+ def custom_forward(*inputs):
+ return module(*inputs)
+
+ return custom_forward
+
+ if is_torch_version(">=", "1.11.0"):
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
+ )
+ else:
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet), hidden_states, temb
+ )
+ else:
+ hidden_states = resnet(hidden_states, temb, scale)
+
+ output_states = output_states + (hidden_states,)
+
+ if self.downsamplers is not None:
+ for downsampler in self.downsamplers:
+ hidden_states = downsampler(hidden_states, temb, scale)
+
+ output_states = output_states + (hidden_states,)
+
+ return hidden_states, output_states
+
+
+class SimpleCrossAttnDownBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ attention_head_dim: int = 1,
+ cross_attention_dim: int = 1280,
+ output_scale_factor: float = 1.0,
+ add_downsample: bool = True,
+ skip_time_act: bool = False,
+ only_cross_attention: bool = False,
+ cross_attention_norm: Optional[str] = None,
+ ):
+ super().__init__()
+
+ self.has_cross_attention = True
+
+ resnets = []
+ attentions = []
+
+ self.attention_head_dim = attention_head_dim
+ self.num_heads = out_channels // self.attention_head_dim
+
+ for i in range(num_layers):
+ in_channels = in_channels if i == 0 else out_channels
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ skip_time_act=skip_time_act,
+ )
+ )
+
+ processor = (
+ AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor()
+ )
+
+ attentions.append(
+ Attention(
+ query_dim=out_channels,
+ cross_attention_dim=out_channels,
+ heads=self.num_heads,
+ dim_head=attention_head_dim,
+ added_kv_proj_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ bias=True,
+ upcast_softmax=True,
+ only_cross_attention=only_cross_attention,
+ cross_attention_norm=cross_attention_norm,
+ processor=processor,
+ )
+ )
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_downsample:
+ self.downsamplers = nn.ModuleList(
+ [
+ ResnetBlock2D(
+ in_channels=out_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ skip_time_act=skip_time_act,
+ down=True,
+ )
+ ]
+ )
+ else:
+ self.downsamplers = None
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
+ ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
+ output_states = ()
+ cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
+
+ lora_scale = cross_attention_kwargs.get("scale", 1.0)
+
+ if attention_mask is None:
+ # if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask.
+ mask = None if encoder_hidden_states is None else encoder_attention_mask
+ else:
+ # when attention_mask is defined: we don't even check for encoder_attention_mask.
+ # this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks.
+ # TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask.
+ # then we can simplify this whole if/else block to:
+ # mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask
+ mask = attention_mask
+
+ for resnet, attn in zip(self.resnets, self.attentions):
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module, return_dict=None):
+ def custom_forward(*inputs):
+ if return_dict is not None:
+ return module(*inputs, return_dict=return_dict)
+ else:
+ return module(*inputs)
+
+ return custom_forward
+
+ hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=mask,
+ **cross_attention_kwargs,
+ )
+ else:
+ hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=mask,
+ **cross_attention_kwargs,
+ )
+
+ output_states = output_states + (hidden_states,)
+
+ if self.downsamplers is not None:
+ for downsampler in self.downsamplers:
+ hidden_states = downsampler(hidden_states, temb, scale=lora_scale)
+
+ output_states = output_states + (hidden_states,)
+
+ return hidden_states, output_states
+
+
+class KDownBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 4,
+ resnet_eps: float = 1e-5,
+ resnet_act_fn: str = "gelu",
+ resnet_group_size: int = 32,
+ add_downsample: bool = False,
+ ):
+ super().__init__()
+ resnets = []
+
+ for i in range(num_layers):
+ in_channels = in_channels if i == 0 else out_channels
+ groups = in_channels // resnet_group_size
+ groups_out = out_channels // resnet_group_size
+
+ resnets.append(
+ ResnetBlockCondNorm2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ dropout=dropout,
+ temb_channels=temb_channels,
+ groups=groups,
+ groups_out=groups_out,
+ eps=resnet_eps,
+ non_linearity=resnet_act_fn,
+ time_embedding_norm="ada_group",
+ conv_shortcut_bias=False,
+ )
+ )
+
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_downsample:
+ # YiYi's comments- might be able to use FirDownsample2D, look into details later
+ self.downsamplers = nn.ModuleList([KDownsample2D()])
+ else:
+ self.downsamplers = None
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0
+ ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
+ output_states = ()
+
+ for resnet in self.resnets:
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module):
+ def custom_forward(*inputs):
+ return module(*inputs)
+
+ return custom_forward
+
+ if is_torch_version(">=", "1.11.0"):
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
+ )
+ else:
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet), hidden_states, temb
+ )
+ else:
+ hidden_states = resnet(hidden_states, temb, scale)
+
+ output_states += (hidden_states,)
+
+ if self.downsamplers is not None:
+ for downsampler in self.downsamplers:
+ hidden_states = downsampler(hidden_states)
+
+ return hidden_states, output_states
+
+
+class KCrossAttnDownBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ cross_attention_dim: int,
+ dropout: float = 0.0,
+ num_layers: int = 4,
+ resnet_group_size: int = 32,
+ add_downsample: bool = True,
+ attention_head_dim: int = 64,
+ add_self_attention: bool = False,
+ resnet_eps: float = 1e-5,
+ resnet_act_fn: str = "gelu",
+ ):
+ super().__init__()
+ resnets = []
+ attentions = []
+
+ self.has_cross_attention = True
+
+ for i in range(num_layers):
+ in_channels = in_channels if i == 0 else out_channels
+ groups = in_channels // resnet_group_size
+ groups_out = out_channels // resnet_group_size
+
+ resnets.append(
+ ResnetBlockCondNorm2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ dropout=dropout,
+ temb_channels=temb_channels,
+ groups=groups,
+ groups_out=groups_out,
+ eps=resnet_eps,
+ non_linearity=resnet_act_fn,
+ time_embedding_norm="ada_group",
+ conv_shortcut_bias=False,
+ )
+ )
+ attentions.append(
+ KAttentionBlock(
+ out_channels,
+ out_channels // attention_head_dim,
+ attention_head_dim,
+ cross_attention_dim=cross_attention_dim,
+ temb_channels=temb_channels,
+ attention_bias=True,
+ add_self_attention=add_self_attention,
+ cross_attention_norm="layer_norm",
+ group_size=resnet_group_size,
+ )
+ )
+
+ self.resnets = nn.ModuleList(resnets)
+ self.attentions = nn.ModuleList(attentions)
+
+ if add_downsample:
+ self.downsamplers = nn.ModuleList([KDownsample2D()])
+ else:
+ self.downsamplers = None
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
+ ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
+ output_states = ()
+ lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+
+ for resnet, attn in zip(self.resnets, self.attentions):
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module, return_dict=None):
+ def custom_forward(*inputs):
+ if return_dict is not None:
+ return module(*inputs, return_dict=return_dict)
+ else:
+ return module(*inputs)
+
+ return custom_forward
+
+ ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet),
+ hidden_states,
+ temb,
+ **ckpt_kwargs,
+ )
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ emb=temb,
+ attention_mask=attention_mask,
+ cross_attention_kwargs=cross_attention_kwargs,
+ encoder_attention_mask=encoder_attention_mask,
+ )
+ else:
+ hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ emb=temb,
+ attention_mask=attention_mask,
+ cross_attention_kwargs=cross_attention_kwargs,
+ encoder_attention_mask=encoder_attention_mask,
+ )
+
+ if self.downsamplers is None:
+ output_states += (None,)
+ else:
+ output_states += (hidden_states,)
+
+ if self.downsamplers is not None:
+ for downsampler in self.downsamplers:
+ hidden_states = downsampler(hidden_states)
+
+ return hidden_states, output_states
+
+
+class AttnUpBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ prev_output_channel: int,
+ out_channels: int,
+ temb_channels: int,
+ resolution_idx: int = None,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ attention_head_dim: int = 1,
+ output_scale_factor: float = 1.0,
+ upsample_type: str = "conv",
+ ):
+ super().__init__()
+ resnets = []
+ attentions = []
+
+ self.upsample_type = upsample_type
+
+ if attention_head_dim is None:
+ logger.warn(
+ f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}."
+ )
+ attention_head_dim = out_channels
+
+ for i in range(num_layers):
+ res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+ resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=resnet_in_channels + res_skip_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+ attentions.append(
+ Attention(
+ out_channels,
+ heads=out_channels // attention_head_dim,
+ dim_head=attention_head_dim,
+ rescale_output_factor=output_scale_factor,
+ eps=resnet_eps,
+ norm_num_groups=resnet_groups,
+ residual_connection=True,
+ bias=True,
+ upcast_softmax=True,
+ _from_deprecated_attn_block=True,
+ )
+ )
+
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ if upsample_type == "conv":
+ self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
+ elif upsample_type == "resnet":
+ self.upsamplers = nn.ModuleList(
+ [
+ ResnetBlock2D(
+ in_channels=out_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ up=True,
+ )
+ ]
+ )
+ else:
+ self.upsamplers = None
+
+ self.resolution_idx = resolution_idx
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+ temb: Optional[torch.FloatTensor] = None,
+ upsample_size: Optional[int] = None,
+ scale: float = 1.0,
+ ) -> torch.FloatTensor:
+ for resnet, attn in zip(self.resnets, self.attentions):
+ # pop res hidden states
+ res_hidden_states = res_hidden_states_tuple[-1]
+ res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+ hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+ hidden_states = resnet(hidden_states, temb, scale=scale)
+ cross_attention_kwargs = {"scale": scale}
+ hidden_states = attn(hidden_states, **cross_attention_kwargs)
+
+ if self.upsamplers is not None:
+ for upsampler in self.upsamplers:
+ if self.upsample_type == "resnet":
+ hidden_states = upsampler(hidden_states, temb=temb, scale=scale)
+ else:
+ hidden_states = upsampler(hidden_states, scale=scale)
+
+ return hidden_states
+
+
+class CrossAttnUpBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ prev_output_channel: int,
+ temb_channels: int,
+ resolution_idx: Optional[int] = None,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ num_attention_heads: int = 1,
+ cross_attention_dim: int = 1280,
+ output_scale_factor: float = 1.0,
+ add_upsample: bool = True,
+ dual_cross_attention: bool = False,
+ use_linear_projection: bool = False,
+ only_cross_attention: bool = False,
+ upcast_attention: bool = False,
+ attention_type: str = "default",
+ ):
+ super().__init__()
+ resnets = []
+ attentions = []
+
+ self.has_cross_attention = True
+ self.num_attention_heads = num_attention_heads
+
+ if isinstance(transformer_layers_per_block, int):
+ transformer_layers_per_block = [transformer_layers_per_block] * num_layers
+
+ for i in range(num_layers):
+ res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+ resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=resnet_in_channels + res_skip_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+ if not dual_cross_attention:
+ attentions.append(
+ Transformer2DModel(
+ num_attention_heads,
+ out_channels // num_attention_heads,
+ in_channels=out_channels,
+ num_layers=transformer_layers_per_block[i],
+ cross_attention_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention,
+ upcast_attention=upcast_attention,
+ attention_type=attention_type,
+ )
+ )
+ else:
+ attentions.append(
+ DualTransformer2DModel(
+ num_attention_heads,
+ out_channels // num_attention_heads,
+ in_channels=out_channels,
+ num_layers=1,
+ cross_attention_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ )
+ )
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_upsample:
+ self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
+ else:
+ self.upsamplers = None
+
+ self.gradient_checkpointing = False
+ self.resolution_idx = resolution_idx
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+ temb: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ upsample_size: Optional[int] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
+ return_res_samples: Optional[bool] = False,
+ up_block_add_samples: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+ is_freeu_enabled = (
+ getattr(self, "s1", None)
+ and getattr(self, "s2", None)
+ and getattr(self, "b1", None)
+ and getattr(self, "b2", None)
+ )
+ if return_res_samples:
+ output_states = ()
+
+ for resnet, attn in zip(self.resnets, self.attentions):
+ # pop res hidden states
+ res_hidden_states = res_hidden_states_tuple[-1]
+ res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+
+ # FreeU: Only operate on the first two stages
+ if is_freeu_enabled:
+ hidden_states, res_hidden_states = apply_freeu(
+ self.resolution_idx,
+ hidden_states,
+ res_hidden_states,
+ s1=self.s1,
+ s2=self.s2,
+ b1=self.b1,
+ b2=self.b2,
+ )
+
+ hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module, return_dict=None):
+ def custom_forward(*inputs):
+ if return_dict is not None:
+ return module(*inputs, return_dict=return_dict)
+ else:
+ return module(*inputs)
+
+ return custom_forward
+
+ ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet),
+ hidden_states,
+ temb,
+ **ckpt_kwargs,
+ )
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ attention_mask=attention_mask,
+ encoder_attention_mask=encoder_attention_mask,
+ return_dict=False,
+ )[0]
+ else:
+ hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ attention_mask=attention_mask,
+ encoder_attention_mask=encoder_attention_mask,
+ return_dict=False,
+ )[0]
+ if return_res_samples:
+ output_states = output_states + (hidden_states,)
+ if up_block_add_samples is not None:
+ hidden_states = hidden_states + up_block_add_samples.pop(0)
+
+ if self.upsamplers is not None:
+ for upsampler in self.upsamplers:
+ hidden_states = upsampler(hidden_states, upsample_size, scale=lora_scale)
+ if return_res_samples:
+ output_states = output_states + (hidden_states,)
+ if up_block_add_samples is not None:
+ hidden_states = hidden_states + up_block_add_samples.pop(0)
+
+ if return_res_samples:
+ return hidden_states, output_states
+ else:
+ return hidden_states
+
+
+class UpBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ prev_output_channel: int,
+ out_channels: int,
+ temb_channels: int,
+ resolution_idx: Optional[int] = None,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ output_scale_factor: float = 1.0,
+ add_upsample: bool = True,
+ ):
+ super().__init__()
+ resnets = []
+
+ for i in range(num_layers):
+ res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+ resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=resnet_in_channels + res_skip_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_upsample:
+ self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
+ else:
+ self.upsamplers = None
+
+ self.gradient_checkpointing = False
+ self.resolution_idx = resolution_idx
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+ temb: Optional[torch.FloatTensor] = None,
+ upsample_size: Optional[int] = None,
+ scale: float = 1.0,
+ return_res_samples: Optional[bool] = False,
+ up_block_add_samples: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ is_freeu_enabled = (
+ getattr(self, "s1", None)
+ and getattr(self, "s2", None)
+ and getattr(self, "b1", None)
+ and getattr(self, "b2", None)
+ )
+ if return_res_samples:
+ output_states = ()
+
+ for resnet in self.resnets:
+ # pop res hidden states
+ res_hidden_states = res_hidden_states_tuple[-1]
+ res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+
+ # FreeU: Only operate on the first two stages
+ if is_freeu_enabled:
+ hidden_states, res_hidden_states = apply_freeu(
+ self.resolution_idx,
+ hidden_states,
+ res_hidden_states,
+ s1=self.s1,
+ s2=self.s2,
+ b1=self.b1,
+ b2=self.b2,
+ )
+
+ hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module):
+ def custom_forward(*inputs):
+ return module(*inputs)
+
+ return custom_forward
+
+ if is_torch_version(">=", "1.11.0"):
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
+ )
+ else:
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet), hidden_states, temb
+ )
+ else:
+ hidden_states = resnet(hidden_states, temb, scale=scale)
+
+ if return_res_samples:
+ output_states = output_states + (hidden_states,)
+ if up_block_add_samples is not None:
+ hidden_states = hidden_states + up_block_add_samples.pop(0) # todo: add before or after
+
+ if self.upsamplers is not None:
+ for upsampler in self.upsamplers:
+ hidden_states = upsampler(hidden_states, upsample_size, scale=scale)
+
+ if return_res_samples:
+ output_states = output_states + (hidden_states,)
+ if up_block_add_samples is not None:
+ hidden_states = hidden_states + up_block_add_samples.pop(0) # todo: add before or after
+
+ if return_res_samples:
+ return hidden_states, output_states
+ else:
+ return hidden_states
+
+
+class UpDecoderBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ resolution_idx: Optional[int] = None,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default", # default, spatial
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ output_scale_factor: float = 1.0,
+ add_upsample: bool = True,
+ temb_channels: Optional[int] = None,
+ ):
+ super().__init__()
+ resnets = []
+
+ for i in range(num_layers):
+ input_channels = in_channels if i == 0 else out_channels
+
+ if resnet_time_scale_shift == "spatial":
+ resnets.append(
+ ResnetBlockCondNorm2D(
+ in_channels=input_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm="spatial",
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ )
+ )
+ else:
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=input_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_upsample:
+ self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
+ else:
+ self.upsamplers = None
+
+ self.resolution_idx = resolution_idx
+
+ def forward(
+ self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0
+ ) -> torch.FloatTensor:
+ for resnet in self.resnets:
+ hidden_states = resnet(hidden_states, temb=temb, scale=scale)
+
+ if self.upsamplers is not None:
+ for upsampler in self.upsamplers:
+ hidden_states = upsampler(hidden_states)
+
+ return hidden_states
+
+
+class AttnUpDecoderBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ resolution_idx: Optional[int] = None,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ attention_head_dim: int = 1,
+ output_scale_factor: float = 1.0,
+ add_upsample: bool = True,
+ temb_channels: Optional[int] = None,
+ ):
+ super().__init__()
+ resnets = []
+ attentions = []
+
+ if attention_head_dim is None:
+ logger.warn(
+ f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `out_channels`: {out_channels}."
+ )
+ attention_head_dim = out_channels
+
+ for i in range(num_layers):
+ input_channels = in_channels if i == 0 else out_channels
+
+ if resnet_time_scale_shift == "spatial":
+ resnets.append(
+ ResnetBlockCondNorm2D(
+ in_channels=input_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm="spatial",
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ )
+ )
+ else:
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=input_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+
+ attentions.append(
+ Attention(
+ out_channels,
+ heads=out_channels // attention_head_dim,
+ dim_head=attention_head_dim,
+ rescale_output_factor=output_scale_factor,
+ eps=resnet_eps,
+ norm_num_groups=resnet_groups if resnet_time_scale_shift != "spatial" else None,
+ spatial_norm_dim=temb_channels if resnet_time_scale_shift == "spatial" else None,
+ residual_connection=True,
+ bias=True,
+ upcast_softmax=True,
+ _from_deprecated_attn_block=True,
+ )
+ )
+
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_upsample:
+ self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
+ else:
+ self.upsamplers = None
+
+ self.resolution_idx = resolution_idx
+
+ def forward(
+ self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0
+ ) -> torch.FloatTensor:
+ for resnet, attn in zip(self.resnets, self.attentions):
+ hidden_states = resnet(hidden_states, temb=temb, scale=scale)
+ cross_attention_kwargs = {"scale": scale}
+ hidden_states = attn(hidden_states, temb=temb, **cross_attention_kwargs)
+
+ if self.upsamplers is not None:
+ for upsampler in self.upsamplers:
+ hidden_states = upsampler(hidden_states, scale=scale)
+
+ return hidden_states
+
+
+class AttnSkipUpBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ prev_output_channel: int,
+ out_channels: int,
+ temb_channels: int,
+ resolution_idx: Optional[int] = None,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_pre_norm: bool = True,
+ attention_head_dim: int = 1,
+ output_scale_factor: float = np.sqrt(2.0),
+ add_upsample: bool = True,
+ ):
+ super().__init__()
+ self.attentions = nn.ModuleList([])
+ self.resnets = nn.ModuleList([])
+
+ for i in range(num_layers):
+ res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+ resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+ self.resnets.append(
+ ResnetBlock2D(
+ in_channels=resnet_in_channels + res_skip_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=min(resnet_in_channels + res_skip_channels // 4, 32),
+ groups_out=min(out_channels // 4, 32),
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+
+ if attention_head_dim is None:
+ logger.warn(
+ f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `out_channels`: {out_channels}."
+ )
+ attention_head_dim = out_channels
+
+ self.attentions.append(
+ Attention(
+ out_channels,
+ heads=out_channels // attention_head_dim,
+ dim_head=attention_head_dim,
+ rescale_output_factor=output_scale_factor,
+ eps=resnet_eps,
+ norm_num_groups=32,
+ residual_connection=True,
+ bias=True,
+ upcast_softmax=True,
+ _from_deprecated_attn_block=True,
+ )
+ )
+
+ self.upsampler = FirUpsample2D(in_channels, out_channels=out_channels)
+ if add_upsample:
+ self.resnet_up = ResnetBlock2D(
+ in_channels=out_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=min(out_channels // 4, 32),
+ groups_out=min(out_channels // 4, 32),
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ use_in_shortcut=True,
+ up=True,
+ kernel="fir",
+ )
+ self.skip_conv = nn.Conv2d(out_channels, 3, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+ self.skip_norm = torch.nn.GroupNorm(
+ num_groups=min(out_channels // 4, 32), num_channels=out_channels, eps=resnet_eps, affine=True
+ )
+ self.act = nn.SiLU()
+ else:
+ self.resnet_up = None
+ self.skip_conv = None
+ self.skip_norm = None
+ self.act = None
+
+ self.resolution_idx = resolution_idx
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+ temb: Optional[torch.FloatTensor] = None,
+ skip_sample=None,
+ scale: float = 1.0,
+ ) -> Tuple[torch.FloatTensor, torch.FloatTensor]:
+ for resnet in self.resnets:
+ # pop res hidden states
+ res_hidden_states = res_hidden_states_tuple[-1]
+ res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+ hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+ hidden_states = resnet(hidden_states, temb, scale=scale)
+
+ cross_attention_kwargs = {"scale": scale}
+ hidden_states = self.attentions[0](hidden_states, **cross_attention_kwargs)
+
+ if skip_sample is not None:
+ skip_sample = self.upsampler(skip_sample)
+ else:
+ skip_sample = 0
+
+ if self.resnet_up is not None:
+ skip_sample_states = self.skip_norm(hidden_states)
+ skip_sample_states = self.act(skip_sample_states)
+ skip_sample_states = self.skip_conv(skip_sample_states)
+
+ skip_sample = skip_sample + skip_sample_states
+
+ hidden_states = self.resnet_up(hidden_states, temb, scale=scale)
+
+ return hidden_states, skip_sample
+
+
+class SkipUpBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ prev_output_channel: int,
+ out_channels: int,
+ temb_channels: int,
+ resolution_idx: Optional[int] = None,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_pre_norm: bool = True,
+ output_scale_factor: float = np.sqrt(2.0),
+ add_upsample: bool = True,
+ upsample_padding: int = 1,
+ ):
+ super().__init__()
+ self.resnets = nn.ModuleList([])
+
+ for i in range(num_layers):
+ res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+ resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+ self.resnets.append(
+ ResnetBlock2D(
+ in_channels=resnet_in_channels + res_skip_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=min((resnet_in_channels + res_skip_channels) // 4, 32),
+ groups_out=min(out_channels // 4, 32),
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+
+ self.upsampler = FirUpsample2D(in_channels, out_channels=out_channels)
+ if add_upsample:
+ self.resnet_up = ResnetBlock2D(
+ in_channels=out_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=min(out_channels // 4, 32),
+ groups_out=min(out_channels // 4, 32),
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ use_in_shortcut=True,
+ up=True,
+ kernel="fir",
+ )
+ self.skip_conv = nn.Conv2d(out_channels, 3, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+ self.skip_norm = torch.nn.GroupNorm(
+ num_groups=min(out_channels // 4, 32), num_channels=out_channels, eps=resnet_eps, affine=True
+ )
+ self.act = nn.SiLU()
+ else:
+ self.resnet_up = None
+ self.skip_conv = None
+ self.skip_norm = None
+ self.act = None
+
+ self.resolution_idx = resolution_idx
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+ temb: Optional[torch.FloatTensor] = None,
+ skip_sample=None,
+ scale: float = 1.0,
+ ) -> Tuple[torch.FloatTensor, torch.FloatTensor]:
+ for resnet in self.resnets:
+ # pop res hidden states
+ res_hidden_states = res_hidden_states_tuple[-1]
+ res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+ hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+ hidden_states = resnet(hidden_states, temb, scale=scale)
+
+ if skip_sample is not None:
+ skip_sample = self.upsampler(skip_sample)
+ else:
+ skip_sample = 0
+
+ if self.resnet_up is not None:
+ skip_sample_states = self.skip_norm(hidden_states)
+ skip_sample_states = self.act(skip_sample_states)
+ skip_sample_states = self.skip_conv(skip_sample_states)
+
+ skip_sample = skip_sample + skip_sample_states
+
+ hidden_states = self.resnet_up(hidden_states, temb, scale=scale)
+
+ return hidden_states, skip_sample
+
+
+class ResnetUpsampleBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ prev_output_channel: int,
+ out_channels: int,
+ temb_channels: int,
+ resolution_idx: Optional[int] = None,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ output_scale_factor: float = 1.0,
+ add_upsample: bool = True,
+ skip_time_act: bool = False,
+ ):
+ super().__init__()
+ resnets = []
+
+ for i in range(num_layers):
+ res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+ resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=resnet_in_channels + res_skip_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ skip_time_act=skip_time_act,
+ )
+ )
+
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_upsample:
+ self.upsamplers = nn.ModuleList(
+ [
+ ResnetBlock2D(
+ in_channels=out_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ skip_time_act=skip_time_act,
+ up=True,
+ )
+ ]
+ )
+ else:
+ self.upsamplers = None
+
+ self.gradient_checkpointing = False
+ self.resolution_idx = resolution_idx
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+ temb: Optional[torch.FloatTensor] = None,
+ upsample_size: Optional[int] = None,
+ scale: float = 1.0,
+ ) -> torch.FloatTensor:
+ for resnet in self.resnets:
+ # pop res hidden states
+ res_hidden_states = res_hidden_states_tuple[-1]
+ res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+ hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module):
+ def custom_forward(*inputs):
+ return module(*inputs)
+
+ return custom_forward
+
+ if is_torch_version(">=", "1.11.0"):
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
+ )
+ else:
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet), hidden_states, temb
+ )
+ else:
+ hidden_states = resnet(hidden_states, temb, scale=scale)
+
+ if self.upsamplers is not None:
+ for upsampler in self.upsamplers:
+ hidden_states = upsampler(hidden_states, temb, scale=scale)
+
+ return hidden_states
+
+
+class SimpleCrossAttnUpBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ prev_output_channel: int,
+ temb_channels: int,
+ resolution_idx: Optional[int] = None,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ attention_head_dim: int = 1,
+ cross_attention_dim: int = 1280,
+ output_scale_factor: float = 1.0,
+ add_upsample: bool = True,
+ skip_time_act: bool = False,
+ only_cross_attention: bool = False,
+ cross_attention_norm: Optional[str] = None,
+ ):
+ super().__init__()
+ resnets = []
+ attentions = []
+
+ self.has_cross_attention = True
+ self.attention_head_dim = attention_head_dim
+
+ self.num_heads = out_channels // self.attention_head_dim
+
+ for i in range(num_layers):
+ res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+ resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=resnet_in_channels + res_skip_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ skip_time_act=skip_time_act,
+ )
+ )
+
+ processor = (
+ AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor()
+ )
+
+ attentions.append(
+ Attention(
+ query_dim=out_channels,
+ cross_attention_dim=out_channels,
+ heads=self.num_heads,
+ dim_head=self.attention_head_dim,
+ added_kv_proj_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ bias=True,
+ upcast_softmax=True,
+ only_cross_attention=only_cross_attention,
+ cross_attention_norm=cross_attention_norm,
+ processor=processor,
+ )
+ )
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_upsample:
+ self.upsamplers = nn.ModuleList(
+ [
+ ResnetBlock2D(
+ in_channels=out_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ skip_time_act=skip_time_act,
+ up=True,
+ )
+ ]
+ )
+ else:
+ self.upsamplers = None
+
+ self.gradient_checkpointing = False
+ self.resolution_idx = resolution_idx
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+ temb: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ upsample_size: Optional[int] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
+
+ lora_scale = cross_attention_kwargs.get("scale", 1.0)
+ if attention_mask is None:
+ # if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask.
+ mask = None if encoder_hidden_states is None else encoder_attention_mask
+ else:
+ # when attention_mask is defined: we don't even check for encoder_attention_mask.
+ # this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks.
+ # TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask.
+ # then we can simplify this whole if/else block to:
+ # mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask
+ mask = attention_mask
+
+ for resnet, attn in zip(self.resnets, self.attentions):
+ # resnet
+ # pop res hidden states
+ res_hidden_states = res_hidden_states_tuple[-1]
+ res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+ hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module, return_dict=None):
+ def custom_forward(*inputs):
+ if return_dict is not None:
+ return module(*inputs, return_dict=return_dict)
+ else:
+ return module(*inputs)
+
+ return custom_forward
+
+ hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=mask,
+ **cross_attention_kwargs,
+ )
+ else:
+ hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=mask,
+ **cross_attention_kwargs,
+ )
+
+ if self.upsamplers is not None:
+ for upsampler in self.upsamplers:
+ hidden_states = upsampler(hidden_states, temb, scale=lora_scale)
+
+ return hidden_states
+
+
+class KUpBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ resolution_idx: int,
+ dropout: float = 0.0,
+ num_layers: int = 5,
+ resnet_eps: float = 1e-5,
+ resnet_act_fn: str = "gelu",
+ resnet_group_size: Optional[int] = 32,
+ add_upsample: bool = True,
+ ):
+ super().__init__()
+ resnets = []
+ k_in_channels = 2 * out_channels
+ k_out_channels = in_channels
+ num_layers = num_layers - 1
+
+ for i in range(num_layers):
+ in_channels = k_in_channels if i == 0 else out_channels
+ groups = in_channels // resnet_group_size
+ groups_out = out_channels // resnet_group_size
+
+ resnets.append(
+ ResnetBlockCondNorm2D(
+ in_channels=in_channels,
+ out_channels=k_out_channels if (i == num_layers - 1) else out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=groups,
+ groups_out=groups_out,
+ dropout=dropout,
+ non_linearity=resnet_act_fn,
+ time_embedding_norm="ada_group",
+ conv_shortcut_bias=False,
+ )
+ )
+
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_upsample:
+ self.upsamplers = nn.ModuleList([KUpsample2D()])
+ else:
+ self.upsamplers = None
+
+ self.gradient_checkpointing = False
+ self.resolution_idx = resolution_idx
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+ temb: Optional[torch.FloatTensor] = None,
+ upsample_size: Optional[int] = None,
+ scale: float = 1.0,
+ ) -> torch.FloatTensor:
+ res_hidden_states_tuple = res_hidden_states_tuple[-1]
+ if res_hidden_states_tuple is not None:
+ hidden_states = torch.cat([hidden_states, res_hidden_states_tuple], dim=1)
+
+ for resnet in self.resnets:
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module):
+ def custom_forward(*inputs):
+ return module(*inputs)
+
+ return custom_forward
+
+ if is_torch_version(">=", "1.11.0"):
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
+ )
+ else:
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet), hidden_states, temb
+ )
+ else:
+ hidden_states = resnet(hidden_states, temb, scale=scale)
+
+ if self.upsamplers is not None:
+ for upsampler in self.upsamplers:
+ hidden_states = upsampler(hidden_states)
+
+ return hidden_states
+
+
+class KCrossAttnUpBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ resolution_idx: int,
+ dropout: float = 0.0,
+ num_layers: int = 4,
+ resnet_eps: float = 1e-5,
+ resnet_act_fn: str = "gelu",
+ resnet_group_size: int = 32,
+ attention_head_dim: int = 1, # attention dim_head
+ cross_attention_dim: int = 768,
+ add_upsample: bool = True,
+ upcast_attention: bool = False,
+ ):
+ super().__init__()
+ resnets = []
+ attentions = []
+
+ is_first_block = in_channels == out_channels == temb_channels
+ is_middle_block = in_channels != out_channels
+ add_self_attention = True if is_first_block else False
+
+ self.has_cross_attention = True
+ self.attention_head_dim = attention_head_dim
+
+ # in_channels, and out_channels for the block (k-unet)
+ k_in_channels = out_channels if is_first_block else 2 * out_channels
+ k_out_channels = in_channels
+
+ num_layers = num_layers - 1
+
+ for i in range(num_layers):
+ in_channels = k_in_channels if i == 0 else out_channels
+ groups = in_channels // resnet_group_size
+ groups_out = out_channels // resnet_group_size
+
+ if is_middle_block and (i == num_layers - 1):
+ conv_2d_out_channels = k_out_channels
+ else:
+ conv_2d_out_channels = None
+
+ resnets.append(
+ ResnetBlockCondNorm2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ conv_2d_out_channels=conv_2d_out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=groups,
+ groups_out=groups_out,
+ dropout=dropout,
+ non_linearity=resnet_act_fn,
+ time_embedding_norm="ada_group",
+ conv_shortcut_bias=False,
+ )
+ )
+ attentions.append(
+ KAttentionBlock(
+ k_out_channels if (i == num_layers - 1) else out_channels,
+ k_out_channels // attention_head_dim
+ if (i == num_layers - 1)
+ else out_channels // attention_head_dim,
+ attention_head_dim,
+ cross_attention_dim=cross_attention_dim,
+ temb_channels=temb_channels,
+ attention_bias=True,
+ add_self_attention=add_self_attention,
+ cross_attention_norm="layer_norm",
+ upcast_attention=upcast_attention,
+ )
+ )
+
+ self.resnets = nn.ModuleList(resnets)
+ self.attentions = nn.ModuleList(attentions)
+
+ if add_upsample:
+ self.upsamplers = nn.ModuleList([KUpsample2D()])
+ else:
+ self.upsamplers = None
+
+ self.gradient_checkpointing = False
+ self.resolution_idx = resolution_idx
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+ temb: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ upsample_size: Optional[int] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ res_hidden_states_tuple = res_hidden_states_tuple[-1]
+ if res_hidden_states_tuple is not None:
+ hidden_states = torch.cat([hidden_states, res_hidden_states_tuple], dim=1)
+
+ lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+ for resnet, attn in zip(self.resnets, self.attentions):
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module, return_dict=None):
+ def custom_forward(*inputs):
+ if return_dict is not None:
+ return module(*inputs, return_dict=return_dict)
+ else:
+ return module(*inputs)
+
+ return custom_forward
+
+ ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet),
+ hidden_states,
+ temb,
+ **ckpt_kwargs,
+ )
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ emb=temb,
+ attention_mask=attention_mask,
+ cross_attention_kwargs=cross_attention_kwargs,
+ encoder_attention_mask=encoder_attention_mask,
+ )
+ else:
+ hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ emb=temb,
+ attention_mask=attention_mask,
+ cross_attention_kwargs=cross_attention_kwargs,
+ encoder_attention_mask=encoder_attention_mask,
+ )
+
+ if self.upsamplers is not None:
+ for upsampler in self.upsamplers:
+ hidden_states = upsampler(hidden_states)
+
+ return hidden_states
+
+
+# can potentially later be renamed to `No-feed-forward` attention
+class KAttentionBlock(nn.Module):
+ r"""
+ A basic Transformer block.
+
+ Parameters:
+ dim (`int`): The number of channels in the input and output.
+ num_attention_heads (`int`): The number of heads to use for multi-head attention.
+ attention_head_dim (`int`): The number of channels in each head.
+ dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+ cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
+ attention_bias (`bool`, *optional*, defaults to `False`):
+ Configure if the attention layers should contain a bias parameter.
+ upcast_attention (`bool`, *optional*, defaults to `False`):
+ Set to `True` to upcast the attention computation to `float32`.
+ temb_channels (`int`, *optional*, defaults to 768):
+ The number of channels in the token embedding.
+ add_self_attention (`bool`, *optional*, defaults to `False`):
+ Set to `True` to add self-attention to the block.
+ cross_attention_norm (`str`, *optional*, defaults to `None`):
+ The type of normalization to use for the cross attention. Can be `None`, `layer_norm`, or `group_norm`.
+ group_size (`int`, *optional*, defaults to 32):
+ The number of groups to separate the channels into for group normalization.
+ """
+
+ def __init__(
+ self,
+ dim: int,
+ num_attention_heads: int,
+ attention_head_dim: int,
+ dropout: float = 0.0,
+ cross_attention_dim: Optional[int] = None,
+ attention_bias: bool = False,
+ upcast_attention: bool = False,
+ temb_channels: int = 768, # for ada_group_norm
+ add_self_attention: bool = False,
+ cross_attention_norm: Optional[str] = None,
+ group_size: int = 32,
+ ):
+ super().__init__()
+ self.add_self_attention = add_self_attention
+
+ # 1. Self-Attn
+ if add_self_attention:
+ self.norm1 = AdaGroupNorm(temb_channels, dim, max(1, dim // group_size))
+ self.attn1 = Attention(
+ query_dim=dim,
+ heads=num_attention_heads,
+ dim_head=attention_head_dim,
+ dropout=dropout,
+ bias=attention_bias,
+ cross_attention_dim=None,
+ cross_attention_norm=None,
+ )
+
+ # 2. Cross-Attn
+ self.norm2 = AdaGroupNorm(temb_channels, dim, max(1, dim // group_size))
+ self.attn2 = Attention(
+ query_dim=dim,
+ cross_attention_dim=cross_attention_dim,
+ heads=num_attention_heads,
+ dim_head=attention_head_dim,
+ dropout=dropout,
+ bias=attention_bias,
+ upcast_attention=upcast_attention,
+ cross_attention_norm=cross_attention_norm,
+ )
+
+ def _to_3d(self, hidden_states: torch.FloatTensor, height: int, weight: int) -> torch.FloatTensor:
+ return hidden_states.permute(0, 2, 3, 1).reshape(hidden_states.shape[0], height * weight, -1)
+
+ def _to_4d(self, hidden_states: torch.FloatTensor, height: int, weight: int) -> torch.FloatTensor:
+ return hidden_states.permute(0, 2, 1).reshape(hidden_states.shape[0], -1, height, weight)
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ # TODO: mark emb as non-optional (self.norm2 requires it).
+ # requires assessing impact of change to positional param interface.
+ emb: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
+
+ # 1. Self-Attention
+ if self.add_self_attention:
+ norm_hidden_states = self.norm1(hidden_states, emb)
+
+ height, weight = norm_hidden_states.shape[2:]
+ norm_hidden_states = self._to_3d(norm_hidden_states, height, weight)
+
+ attn_output = self.attn1(
+ norm_hidden_states,
+ encoder_hidden_states=None,
+ attention_mask=attention_mask,
+ **cross_attention_kwargs,
+ )
+ attn_output = self._to_4d(attn_output, height, weight)
+
+ hidden_states = attn_output + hidden_states
+
+ # 2. Cross-Attention/None
+ norm_hidden_states = self.norm2(hidden_states, emb)
+
+ height, weight = norm_hidden_states.shape[2:]
+ norm_hidden_states = self._to_3d(norm_hidden_states, height, weight)
+ attn_output = self.attn2(
+ norm_hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=attention_mask if encoder_hidden_states is None else encoder_attention_mask,
+ **cross_attention_kwargs,
+ )
+ attn_output = self._to_4d(attn_output, height, weight)
+
+ hidden_states = attn_output + hidden_states
+
+ return hidden_states
diff --git a/iopaint/model/power_paint/v2/unet_2d_condition.py b/iopaint/model/power_paint/v2/unet_2d_condition.py
new file mode 100644
index 0000000..b8bb2f4
--- /dev/null
+++ b/iopaint/model/power_paint/v2/unet_2d_condition.py
@@ -0,0 +1,1543 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.loaders import PeftAdapterMixin, UNet2DConditionLoadersMixin
+from diffusers.utils import (
+ USE_PEFT_BACKEND,
+ BaseOutput,
+ deprecate,
+ logging,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from diffusers.models.activations import get_activation
+from diffusers.models.attention_processor import (
+ ADDED_KV_ATTENTION_PROCESSORS,
+ CROSS_ATTENTION_PROCESSORS,
+ Attention,
+ AttentionProcessor,
+ AttnAddedKVProcessor,
+ AttnProcessor,
+)
+from diffusers.models.embeddings import (
+ GaussianFourierProjection,
+ GLIGENTextBoundingboxProjection,
+ ImageHintTimeEmbedding,
+ ImageProjection,
+ ImageTimeEmbedding,
+ TextImageProjection,
+ TextImageTimeEmbedding,
+ TextTimeEmbedding,
+ TimestepEmbedding,
+ Timesteps,
+)
+from diffusers.models.modeling_utils import ModelMixin
+from .unet_2d_blocks import (
+ get_down_block,
+ get_mid_block,
+ get_up_block,
+)
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+@dataclass
+class UNet2DConditionOutput(BaseOutput):
+ """
+ The output of [`UNet2DConditionModel`].
+
+ Args:
+ sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+ The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model.
+ """
+
+ sample: torch.FloatTensor = None
+
+
+class UNet2DConditionModel(
+ ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin, PeftAdapterMixin
+):
+ r"""
+ A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample
+ shaped output.
+
+ This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+ for all models (such as downloading or saving).
+
+ Parameters:
+ sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
+ Height and width of input/output sample.
+ in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample.
+ out_channels (`int`, *optional*, defaults to 4): Number of channels in the output.
+ center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
+ flip_sin_to_cos (`bool`, *optional*, defaults to `False`):
+ Whether to flip the sin to cos in the time embedding.
+ freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding.
+ down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
+ The tuple of downsample blocks to use.
+ mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`):
+ Block type for middle of UNet, it can be one of `UNetMidBlock2DCrossAttn`, `UNetMidBlock2D`, or
+ `UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped.
+ up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`):
+ The tuple of upsample blocks to use.
+ only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`):
+ Whether to include self-attention in the basic transformer blocks, see
+ [`~models.attention.BasicTransformerBlock`].
+ block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
+ The tuple of output channels for each block.
+ layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
+ downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution.
+ mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block.
+ dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+ act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+ norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
+ If `None`, normalization and activation layers is skipped in post-processing.
+ norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
+ cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
+ The dimension of the cross attention features.
+ transformer_layers_per_block (`int`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1):
+ The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
+ [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+ [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+ reverse_transformer_layers_per_block : (`Tuple[Tuple]`, *optional*, defaults to None):
+ The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`], in the upsampling
+ blocks of the U-Net. Only relevant if `transformer_layers_per_block` is of type `Tuple[Tuple]` and for
+ [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+ [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+ encoder_hid_dim (`int`, *optional*, defaults to None):
+ If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
+ dimension to `cross_attention_dim`.
+ encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
+ If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
+ embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
+ attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
+ num_attention_heads (`int`, *optional*):
+ The number of attention heads. If not defined, defaults to `attention_head_dim`
+ resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
+ for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`.
+ class_embed_type (`str`, *optional*, defaults to `None`):
+ The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
+ `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
+ addition_embed_type (`str`, *optional*, defaults to `None`):
+ Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
+ "text". "text" will use the `TextTimeEmbedding` layer.
+ addition_time_embed_dim: (`int`, *optional*, defaults to `None`):
+ Dimension for the timestep embeddings.
+ num_class_embeds (`int`, *optional*, defaults to `None`):
+ Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
+ class conditioning with `class_embed_type` equal to `None`.
+ time_embedding_type (`str`, *optional*, defaults to `positional`):
+ The type of position embedding to use for timesteps. Choose from `positional` or `fourier`.
+ time_embedding_dim (`int`, *optional*, defaults to `None`):
+ An optional override for the dimension of the projected time embedding.
+ time_embedding_act_fn (`str`, *optional*, defaults to `None`):
+ Optional activation function to use only once on the time embeddings before they are passed to the rest of
+ the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`.
+ timestep_post_act (`str`, *optional*, defaults to `None`):
+ The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`.
+ time_cond_proj_dim (`int`, *optional*, defaults to `None`):
+ The dimension of `cond_proj` layer in the timestep embedding.
+ conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer. conv_out_kernel (`int`,
+ *optional*, default to `3`): The kernel size of `conv_out` layer. projection_class_embeddings_input_dim (`int`,
+ *optional*): The dimension of the `class_labels` input when
+ `class_embed_type="projection"`. Required when `class_embed_type="projection"`.
+ class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time
+ embeddings with the class embeddings.
+ mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`):
+ Whether to use cross attention with the mid block when using the `UNetMidBlock2DSimpleCrossAttn`. If
+ `only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the
+ `only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False`
+ otherwise.
+ """
+
+ _supports_gradient_checkpointing = True
+
+ @register_to_config
+ def __init__(
+ self,
+ sample_size: Optional[int] = None,
+ in_channels: int = 4,
+ out_channels: int = 4,
+ center_input_sample: bool = False,
+ flip_sin_to_cos: bool = True,
+ freq_shift: int = 0,
+ down_block_types: Tuple[str] = (
+ "CrossAttnDownBlock2D",
+ "CrossAttnDownBlock2D",
+ "CrossAttnDownBlock2D",
+ "DownBlock2D",
+ ),
+ mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
+ up_block_types: Tuple[str] = (
+ "UpBlock2D",
+ "CrossAttnUpBlock2D",
+ "CrossAttnUpBlock2D",
+ "CrossAttnUpBlock2D",
+ ),
+ only_cross_attention: Union[bool, Tuple[bool]] = False,
+ block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+ layers_per_block: Union[int, Tuple[int]] = 2,
+ downsample_padding: int = 1,
+ mid_block_scale_factor: float = 1,
+ dropout: float = 0.0,
+ act_fn: str = "silu",
+ norm_num_groups: Optional[int] = 32,
+ norm_eps: float = 1e-5,
+ cross_attention_dim: Union[int, Tuple[int]] = 1280,
+ transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1,
+ reverse_transformer_layers_per_block: Optional[Tuple[Tuple[int]]] = None,
+ encoder_hid_dim: Optional[int] = None,
+ encoder_hid_dim_type: Optional[str] = None,
+ attention_head_dim: Union[int, Tuple[int]] = 8,
+ num_attention_heads: Optional[Union[int, Tuple[int]]] = None,
+ dual_cross_attention: bool = False,
+ use_linear_projection: bool = False,
+ class_embed_type: Optional[str] = None,
+ addition_embed_type: Optional[str] = None,
+ addition_time_embed_dim: Optional[int] = None,
+ num_class_embeds: Optional[int] = None,
+ upcast_attention: bool = False,
+ resnet_time_scale_shift: str = "default",
+ resnet_skip_time_act: bool = False,
+ resnet_out_scale_factor: float = 1.0,
+ time_embedding_type: str = "positional",
+ time_embedding_dim: Optional[int] = None,
+ time_embedding_act_fn: Optional[str] = None,
+ timestep_post_act: Optional[str] = None,
+ time_cond_proj_dim: Optional[int] = None,
+ conv_in_kernel: int = 3,
+ conv_out_kernel: int = 3,
+ projection_class_embeddings_input_dim: Optional[int] = None,
+ attention_type: str = "default",
+ class_embeddings_concat: bool = False,
+ mid_block_only_cross_attention: Optional[bool] = None,
+ cross_attention_norm: Optional[str] = None,
+ addition_embed_type_num_heads: int = 64,
+ ):
+ super().__init__()
+
+ self.sample_size = sample_size
+
+ if num_attention_heads is not None:
+ raise ValueError(
+ "At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19."
+ )
+
+ # If `num_attention_heads` is not defined (which is the case for most models)
+ # it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
+ # The reason for this behavior is to correct for incorrectly named variables that were introduced
+ # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
+ # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
+ # which is why we correct for the naming here.
+ num_attention_heads = num_attention_heads or attention_head_dim
+
+ # Check inputs
+ self._check_config(
+ down_block_types=down_block_types,
+ up_block_types=up_block_types,
+ only_cross_attention=only_cross_attention,
+ block_out_channels=block_out_channels,
+ layers_per_block=layers_per_block,
+ cross_attention_dim=cross_attention_dim,
+ transformer_layers_per_block=transformer_layers_per_block,
+ reverse_transformer_layers_per_block=reverse_transformer_layers_per_block,
+ attention_head_dim=attention_head_dim,
+ num_attention_heads=num_attention_heads,
+ )
+
+ # input
+ conv_in_padding = (conv_in_kernel - 1) // 2
+ self.conv_in = nn.Conv2d(
+ in_channels,
+ block_out_channels[0],
+ kernel_size=conv_in_kernel,
+ padding=conv_in_padding,
+ )
+
+ # time
+ time_embed_dim, timestep_input_dim = self._set_time_proj(
+ time_embedding_type,
+ block_out_channels=block_out_channels,
+ flip_sin_to_cos=flip_sin_to_cos,
+ freq_shift=freq_shift,
+ time_embedding_dim=time_embedding_dim,
+ )
+
+ self.time_embedding = TimestepEmbedding(
+ timestep_input_dim,
+ time_embed_dim,
+ act_fn=act_fn,
+ post_act_fn=timestep_post_act,
+ cond_proj_dim=time_cond_proj_dim,
+ )
+
+ self._set_encoder_hid_proj(
+ encoder_hid_dim_type,
+ cross_attention_dim=cross_attention_dim,
+ encoder_hid_dim=encoder_hid_dim,
+ )
+
+ # class embedding
+ self._set_class_embedding(
+ class_embed_type,
+ act_fn=act_fn,
+ num_class_embeds=num_class_embeds,
+ projection_class_embeddings_input_dim=projection_class_embeddings_input_dim,
+ time_embed_dim=time_embed_dim,
+ timestep_input_dim=timestep_input_dim,
+ )
+
+ self._set_add_embedding(
+ addition_embed_type,
+ addition_embed_type_num_heads=addition_embed_type_num_heads,
+ addition_time_embed_dim=addition_time_embed_dim,
+ cross_attention_dim=cross_attention_dim,
+ encoder_hid_dim=encoder_hid_dim,
+ flip_sin_to_cos=flip_sin_to_cos,
+ freq_shift=freq_shift,
+ projection_class_embeddings_input_dim=projection_class_embeddings_input_dim,
+ time_embed_dim=time_embed_dim,
+ )
+
+ if time_embedding_act_fn is None:
+ self.time_embed_act = None
+ else:
+ self.time_embed_act = get_activation(time_embedding_act_fn)
+
+ self.down_blocks = nn.ModuleList([])
+ self.up_blocks = nn.ModuleList([])
+
+ if isinstance(only_cross_attention, bool):
+ if mid_block_only_cross_attention is None:
+ mid_block_only_cross_attention = only_cross_attention
+
+ only_cross_attention = [only_cross_attention] * len(down_block_types)
+
+ if mid_block_only_cross_attention is None:
+ mid_block_only_cross_attention = False
+
+ if isinstance(num_attention_heads, int):
+ num_attention_heads = (num_attention_heads,) * len(down_block_types)
+
+ if isinstance(attention_head_dim, int):
+ attention_head_dim = (attention_head_dim,) * len(down_block_types)
+
+ if isinstance(cross_attention_dim, int):
+ cross_attention_dim = (cross_attention_dim,) * len(down_block_types)
+
+ if isinstance(layers_per_block, int):
+ layers_per_block = [layers_per_block] * len(down_block_types)
+
+ if isinstance(transformer_layers_per_block, int):
+ transformer_layers_per_block = [transformer_layers_per_block] * len(
+ down_block_types
+ )
+
+ if class_embeddings_concat:
+ # The time embeddings are concatenated with the class embeddings. The dimension of the
+ # time embeddings passed to the down, middle, and up blocks is twice the dimension of the
+ # regular time embeddings
+ blocks_time_embed_dim = time_embed_dim * 2
+ else:
+ blocks_time_embed_dim = time_embed_dim
+
+ # down
+ output_channel = block_out_channels[0]
+ for i, down_block_type in enumerate(down_block_types):
+ input_channel = output_channel
+ output_channel = block_out_channels[i]
+ is_final_block = i == len(block_out_channels) - 1
+
+ down_block = get_down_block(
+ down_block_type,
+ num_layers=layers_per_block[i],
+ transformer_layers_per_block=transformer_layers_per_block[i],
+ in_channels=input_channel,
+ out_channels=output_channel,
+ temb_channels=blocks_time_embed_dim,
+ add_downsample=not is_final_block,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ resnet_groups=norm_num_groups,
+ cross_attention_dim=cross_attention_dim[i],
+ num_attention_heads=num_attention_heads[i],
+ downsample_padding=downsample_padding,
+ dual_cross_attention=dual_cross_attention,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention[i],
+ upcast_attention=upcast_attention,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ attention_type=attention_type,
+ resnet_skip_time_act=resnet_skip_time_act,
+ resnet_out_scale_factor=resnet_out_scale_factor,
+ cross_attention_norm=cross_attention_norm,
+ attention_head_dim=attention_head_dim[i]
+ if attention_head_dim[i] is not None
+ else output_channel,
+ dropout=dropout,
+ )
+ self.down_blocks.append(down_block)
+
+ # mid
+ self.mid_block = get_mid_block(
+ mid_block_type,
+ temb_channels=blocks_time_embed_dim,
+ in_channels=block_out_channels[-1],
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ resnet_groups=norm_num_groups,
+ output_scale_factor=mid_block_scale_factor,
+ transformer_layers_per_block=transformer_layers_per_block[-1],
+ num_attention_heads=num_attention_heads[-1],
+ cross_attention_dim=cross_attention_dim[-1],
+ dual_cross_attention=dual_cross_attention,
+ use_linear_projection=use_linear_projection,
+ mid_block_only_cross_attention=mid_block_only_cross_attention,
+ upcast_attention=upcast_attention,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ attention_type=attention_type,
+ resnet_skip_time_act=resnet_skip_time_act,
+ cross_attention_norm=cross_attention_norm,
+ attention_head_dim=attention_head_dim[-1],
+ dropout=dropout,
+ )
+
+ # count how many layers upsample the images
+ self.num_upsamplers = 0
+
+ # up
+ reversed_block_out_channels = list(reversed(block_out_channels))
+ reversed_num_attention_heads = list(reversed(num_attention_heads))
+ reversed_layers_per_block = list(reversed(layers_per_block))
+ reversed_cross_attention_dim = list(reversed(cross_attention_dim))
+ reversed_transformer_layers_per_block = (
+ list(reversed(transformer_layers_per_block))
+ if reverse_transformer_layers_per_block is None
+ else reverse_transformer_layers_per_block
+ )
+ only_cross_attention = list(reversed(only_cross_attention))
+
+ output_channel = reversed_block_out_channels[0]
+ for i, up_block_type in enumerate(up_block_types):
+ is_final_block = i == len(block_out_channels) - 1
+
+ prev_output_channel = output_channel
+ output_channel = reversed_block_out_channels[i]
+ input_channel = reversed_block_out_channels[
+ min(i + 1, len(block_out_channels) - 1)
+ ]
+
+ # add upsample block for all BUT final layer
+ if not is_final_block:
+ add_upsample = True
+ self.num_upsamplers += 1
+ else:
+ add_upsample = False
+
+ up_block = get_up_block(
+ up_block_type,
+ num_layers=reversed_layers_per_block[i] + 1,
+ transformer_layers_per_block=reversed_transformer_layers_per_block[i],
+ in_channels=input_channel,
+ out_channels=output_channel,
+ prev_output_channel=prev_output_channel,
+ temb_channels=blocks_time_embed_dim,
+ add_upsample=add_upsample,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ resolution_idx=i,
+ resnet_groups=norm_num_groups,
+ cross_attention_dim=reversed_cross_attention_dim[i],
+ num_attention_heads=reversed_num_attention_heads[i],
+ dual_cross_attention=dual_cross_attention,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention[i],
+ upcast_attention=upcast_attention,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ attention_type=attention_type,
+ resnet_skip_time_act=resnet_skip_time_act,
+ resnet_out_scale_factor=resnet_out_scale_factor,
+ cross_attention_norm=cross_attention_norm,
+ attention_head_dim=attention_head_dim[i]
+ if attention_head_dim[i] is not None
+ else output_channel,
+ dropout=dropout,
+ )
+ self.up_blocks.append(up_block)
+ prev_output_channel = output_channel
+
+ # out
+ if norm_num_groups is not None:
+ self.conv_norm_out = nn.GroupNorm(
+ num_channels=block_out_channels[0],
+ num_groups=norm_num_groups,
+ eps=norm_eps,
+ )
+
+ self.conv_act = get_activation(act_fn)
+
+ else:
+ self.conv_norm_out = None
+ self.conv_act = None
+
+ conv_out_padding = (conv_out_kernel - 1) // 2
+ self.conv_out = nn.Conv2d(
+ block_out_channels[0],
+ out_channels,
+ kernel_size=conv_out_kernel,
+ padding=conv_out_padding,
+ )
+
+ self._set_pos_net_if_use_gligen(
+ attention_type=attention_type, cross_attention_dim=cross_attention_dim
+ )
+
+ def _check_config(
+ self,
+ down_block_types: Tuple[str],
+ up_block_types: Tuple[str],
+ only_cross_attention: Union[bool, Tuple[bool]],
+ block_out_channels: Tuple[int],
+ layers_per_block: Union[int, Tuple[int]],
+ cross_attention_dim: Union[int, Tuple[int]],
+ transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple[int]]],
+ reverse_transformer_layers_per_block: bool,
+ attention_head_dim: int,
+ num_attention_heads: Optional[Union[int, Tuple[int]]],
+ ):
+ if len(down_block_types) != len(up_block_types):
+ raise ValueError(
+ f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
+ )
+
+ if len(block_out_channels) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
+ )
+
+ if not isinstance(only_cross_attention, bool) and len(
+ only_cross_attention
+ ) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
+ )
+
+ if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(
+ down_block_types
+ ):
+ raise ValueError(
+ f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
+ )
+
+ if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(
+ down_block_types
+ ):
+ raise ValueError(
+ f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}."
+ )
+
+ if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(
+ down_block_types
+ ):
+ raise ValueError(
+ f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}."
+ )
+
+ if not isinstance(layers_per_block, int) and len(layers_per_block) != len(
+ down_block_types
+ ):
+ raise ValueError(
+ f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}."
+ )
+ if (
+ isinstance(transformer_layers_per_block, list)
+ and reverse_transformer_layers_per_block is None
+ ):
+ for layer_number_per_block in transformer_layers_per_block:
+ if isinstance(layer_number_per_block, list):
+ raise ValueError(
+ "Must provide 'reverse_transformer_layers_per_block` if using asymmetrical UNet."
+ )
+
+ def _set_time_proj(
+ self,
+ time_embedding_type: str,
+ block_out_channels: int,
+ flip_sin_to_cos: bool,
+ freq_shift: float,
+ time_embedding_dim: int,
+ ) -> Tuple[int, int]:
+ if time_embedding_type == "fourier":
+ time_embed_dim = time_embedding_dim or block_out_channels[0] * 2
+ if time_embed_dim % 2 != 0:
+ raise ValueError(
+ f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}."
+ )
+ self.time_proj = GaussianFourierProjection(
+ time_embed_dim // 2,
+ set_W_to_weight=False,
+ log=False,
+ flip_sin_to_cos=flip_sin_to_cos,
+ )
+ timestep_input_dim = time_embed_dim
+ elif time_embedding_type == "positional":
+ time_embed_dim = time_embedding_dim or block_out_channels[0] * 4
+
+ self.time_proj = Timesteps(
+ block_out_channels[0], flip_sin_to_cos, freq_shift
+ )
+ timestep_input_dim = block_out_channels[0]
+ else:
+ raise ValueError(
+ f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`."
+ )
+
+ return time_embed_dim, timestep_input_dim
+
+ def _set_encoder_hid_proj(
+ self,
+ encoder_hid_dim_type: Optional[str],
+ cross_attention_dim: Union[int, Tuple[int]],
+ encoder_hid_dim: Optional[int],
+ ):
+ if encoder_hid_dim_type is None and encoder_hid_dim is not None:
+ encoder_hid_dim_type = "text_proj"
+ self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
+ logger.info(
+ "encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined."
+ )
+
+ if encoder_hid_dim is None and encoder_hid_dim_type is not None:
+ raise ValueError(
+ f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}."
+ )
+
+ if encoder_hid_dim_type == "text_proj":
+ self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim)
+ elif encoder_hid_dim_type == "text_image_proj":
+ # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+ # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+ # case when `addition_embed_type == "text_image_proj"` (Kadinsky 2.1)`
+ self.encoder_hid_proj = TextImageProjection(
+ text_embed_dim=encoder_hid_dim,
+ image_embed_dim=cross_attention_dim,
+ cross_attention_dim=cross_attention_dim,
+ )
+ elif encoder_hid_dim_type == "image_proj":
+ # Kandinsky 2.2
+ self.encoder_hid_proj = ImageProjection(
+ image_embed_dim=encoder_hid_dim,
+ cross_attention_dim=cross_attention_dim,
+ )
+ elif encoder_hid_dim_type is not None:
+ raise ValueError(
+ f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'."
+ )
+ else:
+ self.encoder_hid_proj = None
+
+ def _set_class_embedding(
+ self,
+ class_embed_type: Optional[str],
+ act_fn: str,
+ num_class_embeds: Optional[int],
+ projection_class_embeddings_input_dim: Optional[int],
+ time_embed_dim: int,
+ timestep_input_dim: int,
+ ):
+ if class_embed_type is None and num_class_embeds is not None:
+ self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
+ elif class_embed_type == "timestep":
+ self.class_embedding = TimestepEmbedding(
+ timestep_input_dim, time_embed_dim, act_fn=act_fn
+ )
+ elif class_embed_type == "identity":
+ self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
+ elif class_embed_type == "projection":
+ if projection_class_embeddings_input_dim is None:
+ raise ValueError(
+ "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
+ )
+ # The projection `class_embed_type` is the same as the timestep `class_embed_type` except
+ # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
+ # 2. it projects from an arbitrary input dimension.
+ #
+ # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
+ # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
+ # As a result, `TimestepEmbedding` can be passed arbitrary vectors.
+ self.class_embedding = TimestepEmbedding(
+ projection_class_embeddings_input_dim, time_embed_dim
+ )
+ elif class_embed_type == "simple_projection":
+ if projection_class_embeddings_input_dim is None:
+ raise ValueError(
+ "`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set"
+ )
+ self.class_embedding = nn.Linear(
+ projection_class_embeddings_input_dim, time_embed_dim
+ )
+ else:
+ self.class_embedding = None
+
+ def _set_add_embedding(
+ self,
+ addition_embed_type: str,
+ addition_embed_type_num_heads: int,
+ addition_time_embed_dim: Optional[int],
+ flip_sin_to_cos: bool,
+ freq_shift: float,
+ cross_attention_dim: Optional[int],
+ encoder_hid_dim: Optional[int],
+ projection_class_embeddings_input_dim: Optional[int],
+ time_embed_dim: int,
+ ):
+ if addition_embed_type == "text":
+ if encoder_hid_dim is not None:
+ text_time_embedding_from_dim = encoder_hid_dim
+ else:
+ text_time_embedding_from_dim = cross_attention_dim
+
+ self.add_embedding = TextTimeEmbedding(
+ text_time_embedding_from_dim,
+ time_embed_dim,
+ num_heads=addition_embed_type_num_heads,
+ )
+ elif addition_embed_type == "text_image":
+ # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+ # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+ # case when `addition_embed_type == "text_image"` (Kadinsky 2.1)`
+ self.add_embedding = TextImageTimeEmbedding(
+ text_embed_dim=cross_attention_dim,
+ image_embed_dim=cross_attention_dim,
+ time_embed_dim=time_embed_dim,
+ )
+ elif addition_embed_type == "text_time":
+ self.add_time_proj = Timesteps(
+ addition_time_embed_dim, flip_sin_to_cos, freq_shift
+ )
+ self.add_embedding = TimestepEmbedding(
+ projection_class_embeddings_input_dim, time_embed_dim
+ )
+ elif addition_embed_type == "image":
+ # Kandinsky 2.2
+ self.add_embedding = ImageTimeEmbedding(
+ image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim
+ )
+ elif addition_embed_type == "image_hint":
+ # Kandinsky 2.2 ControlNet
+ self.add_embedding = ImageHintTimeEmbedding(
+ image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim
+ )
+ elif addition_embed_type is not None:
+ raise ValueError(
+ f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'."
+ )
+
+ def _set_pos_net_if_use_gligen(self, attention_type: str, cross_attention_dim: int):
+ if attention_type in ["gated", "gated-text-image"]:
+ positive_len = 768
+ if isinstance(cross_attention_dim, int):
+ positive_len = cross_attention_dim
+ elif isinstance(cross_attention_dim, tuple) or isinstance(
+ cross_attention_dim, list
+ ):
+ positive_len = cross_attention_dim[0]
+
+ feature_type = "text-only" if attention_type == "gated" else "text-image"
+ self.position_net = GLIGENTextBoundingboxProjection(
+ positive_len=positive_len,
+ out_dim=cross_attention_dim,
+ feature_type=feature_type,
+ )
+
+ @property
+ def attn_processors(self) -> Dict[str, AttentionProcessor]:
+ r"""
+ Returns:
+ `dict` of attention processors: A dictionary containing all attention processors used in the model with
+ indexed by its weight name.
+ """
+ # set recursively
+ processors = {}
+
+ def fn_recursive_add_processors(
+ name: str,
+ module: torch.nn.Module,
+ processors: Dict[str, AttentionProcessor],
+ ):
+ if hasattr(module, "get_processor"):
+ processors[f"{name}.processor"] = module.get_processor(
+ return_deprecated_lora=True
+ )
+
+ for sub_name, child in module.named_children():
+ fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+ return processors
+
+ for name, module in self.named_children():
+ fn_recursive_add_processors(name, module, processors)
+
+ return processors
+
+ def set_attn_processor(
+ self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]
+ ):
+ r"""
+ Sets the attention processor to use to compute attention.
+
+ Parameters:
+ processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+ The instantiated processor class or a dictionary of processor classes that will be set as the processor
+ for **all** `Attention` layers.
+
+ If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+ processor. This is strongly recommended when setting trainable attention processors.
+
+ """
+ count = len(self.attn_processors.keys())
+
+ if isinstance(processor, dict) and len(processor) != count:
+ raise ValueError(
+ f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+ f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+ )
+
+ def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+ if hasattr(module, "set_processor"):
+ if not isinstance(processor, dict):
+ module.set_processor(processor)
+ else:
+ module.set_processor(processor.pop(f"{name}.processor"))
+
+ for sub_name, child in module.named_children():
+ fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+ for name, module in self.named_children():
+ fn_recursive_attn_processor(name, module, processor)
+
+ def set_default_attn_processor(self):
+ """
+ Disables custom attention processors and sets the default attention implementation.
+ """
+ if all(
+ proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS
+ for proc in self.attn_processors.values()
+ ):
+ processor = AttnAddedKVProcessor()
+ elif all(
+ proc.__class__ in CROSS_ATTENTION_PROCESSORS
+ for proc in self.attn_processors.values()
+ ):
+ processor = AttnProcessor()
+ else:
+ raise ValueError(
+ f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+ )
+
+ self.set_attn_processor(processor)
+
+ def set_attention_slice(self, slice_size: Union[str, int, List[int]] = "auto"):
+ r"""
+ Enable sliced attention computation.
+
+ When this option is enabled, the attention module splits the input tensor in slices to compute attention in
+ several steps. This is useful for saving some memory in exchange for a small decrease in speed.
+
+ Args:
+ slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+ When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
+ `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
+ provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+ must be a multiple of `slice_size`.
+ """
+ sliceable_head_dims = []
+
+ def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
+ if hasattr(module, "set_attention_slice"):
+ sliceable_head_dims.append(module.sliceable_head_dim)
+
+ for child in module.children():
+ fn_recursive_retrieve_sliceable_dims(child)
+
+ # retrieve number of attention layers
+ for module in self.children():
+ fn_recursive_retrieve_sliceable_dims(module)
+
+ num_sliceable_layers = len(sliceable_head_dims)
+
+ if slice_size == "auto":
+ # half the attention head size is usually a good trade-off between
+ # speed and memory
+ slice_size = [dim // 2 for dim in sliceable_head_dims]
+ elif slice_size == "max":
+ # make smallest slice possible
+ slice_size = num_sliceable_layers * [1]
+
+ slice_size = (
+ num_sliceable_layers * [slice_size]
+ if not isinstance(slice_size, list)
+ else slice_size
+ )
+
+ if len(slice_size) != len(sliceable_head_dims):
+ raise ValueError(
+ f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
+ f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
+ )
+
+ for i in range(len(slice_size)):
+ size = slice_size[i]
+ dim = sliceable_head_dims[i]
+ if size is not None and size > dim:
+ raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
+
+ # Recursively walk through all the children.
+ # Any children which exposes the set_attention_slice method
+ # gets the message
+ def fn_recursive_set_attention_slice(
+ module: torch.nn.Module, slice_size: List[int]
+ ):
+ if hasattr(module, "set_attention_slice"):
+ module.set_attention_slice(slice_size.pop())
+
+ for child in module.children():
+ fn_recursive_set_attention_slice(child, slice_size)
+
+ reversed_slice_size = list(reversed(slice_size))
+ for module in self.children():
+ fn_recursive_set_attention_slice(module, reversed_slice_size)
+
+ def _set_gradient_checkpointing(self, module, value=False):
+ if hasattr(module, "gradient_checkpointing"):
+ module.gradient_checkpointing = value
+
+ def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
+ r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497.
+
+ The suffixes after the scaling factors represent the stage blocks where they are being applied.
+
+ Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that
+ are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+ Args:
+ s1 (`float`):
+ Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+ mitigate the "oversmoothing effect" in the enhanced denoising process.
+ s2 (`float`):
+ Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+ mitigate the "oversmoothing effect" in the enhanced denoising process.
+ b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+ b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+ """
+ for i, upsample_block in enumerate(self.up_blocks):
+ setattr(upsample_block, "s1", s1)
+ setattr(upsample_block, "s2", s2)
+ setattr(upsample_block, "b1", b1)
+ setattr(upsample_block, "b2", b2)
+
+ def disable_freeu(self):
+ """Disables the FreeU mechanism."""
+ freeu_keys = {"s1", "s2", "b1", "b2"}
+ for i, upsample_block in enumerate(self.up_blocks):
+ for k in freeu_keys:
+ if (
+ hasattr(upsample_block, k)
+ or getattr(upsample_block, k, None) is not None
+ ):
+ setattr(upsample_block, k, None)
+
+ def fuse_qkv_projections(self):
+ """
+ Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query,
+ key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
+
+
+
+ This API is 🧪 experimental.
+
+
+ """
+ self.original_attn_processors = None
+
+ for _, attn_processor in self.attn_processors.items():
+ if "Added" in str(attn_processor.__class__.__name__):
+ raise ValueError(
+ "`fuse_qkv_projections()` is not supported for models having added KV projections."
+ )
+
+ self.original_attn_processors = self.attn_processors
+
+ for module in self.modules():
+ if isinstance(module, Attention):
+ module.fuse_projections(fuse=True)
+
+ def unfuse_qkv_projections(self):
+ """Disables the fused QKV projection if enabled.
+
+
+
+ This API is 🧪 experimental.
+
+
+
+ """
+ if self.original_attn_processors is not None:
+ self.set_attn_processor(self.original_attn_processors)
+
+ def unload_lora(self):
+ """Unloads LoRA weights."""
+ deprecate(
+ "unload_lora",
+ "0.28.0",
+ "Calling `unload_lora()` is deprecated and will be removed in a future version. Please install `peft` and then call `disable_adapters().",
+ )
+ for module in self.modules():
+ if hasattr(module, "set_lora_layer"):
+ module.set_lora_layer(None)
+
+ def get_time_embed(
+ self, sample: torch.Tensor, timestep: Union[torch.Tensor, float, int]
+ ) -> Optional[torch.Tensor]:
+ timesteps = timestep
+ if not torch.is_tensor(timesteps):
+ # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+ # This would be a good case for the `match` statement (Python 3.10+)
+ is_mps = sample.device.type == "mps"
+ if isinstance(timestep, float):
+ dtype = torch.float32 if is_mps else torch.float64
+ else:
+ dtype = torch.int32 if is_mps else torch.int64
+ timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+ elif len(timesteps.shape) == 0:
+ timesteps = timesteps[None].to(sample.device)
+
+ # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+ timesteps = timesteps.expand(sample.shape[0])
+
+ t_emb = self.time_proj(timesteps)
+ # `Timesteps` does not contain any weights and will always return f32 tensors
+ # but time_embedding might actually be running in fp16. so we need to cast here.
+ # there might be better ways to encapsulate this.
+ t_emb = t_emb.to(dtype=sample.dtype)
+ return t_emb
+
+ def get_class_embed(
+ self, sample: torch.Tensor, class_labels: Optional[torch.Tensor]
+ ) -> Optional[torch.Tensor]:
+ class_emb = None
+ if self.class_embedding is not None:
+ if class_labels is None:
+ raise ValueError(
+ "class_labels should be provided when num_class_embeds > 0"
+ )
+
+ if self.config.class_embed_type == "timestep":
+ class_labels = self.time_proj(class_labels)
+
+ # `Timesteps` does not contain any weights and will always return f32 tensors
+ # there might be better ways to encapsulate this.
+ class_labels = class_labels.to(dtype=sample.dtype)
+
+ class_emb = self.class_embedding(class_labels).to(dtype=sample.dtype)
+ return class_emb
+
+ def get_aug_embed(
+ self,
+ emb: torch.Tensor,
+ encoder_hidden_states: torch.Tensor,
+ added_cond_kwargs: Dict[str, Any],
+ ) -> Optional[torch.Tensor]:
+ aug_emb = None
+ if self.config.addition_embed_type == "text":
+ aug_emb = self.add_embedding(encoder_hidden_states)
+ elif self.config.addition_embed_type == "text_image":
+ # Kandinsky 2.1 - style
+ if "image_embeds" not in added_cond_kwargs:
+ raise ValueError(
+ f"{self.__class__} has the config param `addition_embed_type` set to 'text_image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+ )
+
+ image_embs = added_cond_kwargs.get("image_embeds")
+ text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states)
+ aug_emb = self.add_embedding(text_embs, image_embs)
+ elif self.config.addition_embed_type == "text_time":
+ # SDXL - style
+ if "text_embeds" not in added_cond_kwargs:
+ raise ValueError(
+ f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
+ )
+ text_embeds = added_cond_kwargs.get("text_embeds")
+ if "time_ids" not in added_cond_kwargs:
+ raise ValueError(
+ f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
+ )
+ time_ids = added_cond_kwargs.get("time_ids")
+ time_embeds = self.add_time_proj(time_ids.flatten())
+ time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
+ add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
+ add_embeds = add_embeds.to(emb.dtype)
+ aug_emb = self.add_embedding(add_embeds)
+ elif self.config.addition_embed_type == "image":
+ # Kandinsky 2.2 - style
+ if "image_embeds" not in added_cond_kwargs:
+ raise ValueError(
+ f"{self.__class__} has the config param `addition_embed_type` set to 'image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+ )
+ image_embs = added_cond_kwargs.get("image_embeds")
+ aug_emb = self.add_embedding(image_embs)
+ elif self.config.addition_embed_type == "image_hint":
+ # Kandinsky 2.2 - style
+ if (
+ "image_embeds" not in added_cond_kwargs
+ or "hint" not in added_cond_kwargs
+ ):
+ raise ValueError(
+ f"{self.__class__} has the config param `addition_embed_type` set to 'image_hint' which requires the keyword arguments `image_embeds` and `hint` to be passed in `added_cond_kwargs`"
+ )
+ image_embs = added_cond_kwargs.get("image_embeds")
+ hint = added_cond_kwargs.get("hint")
+ aug_emb = self.add_embedding(image_embs, hint)
+ return aug_emb
+
+ def process_encoder_hidden_states(
+ self, encoder_hidden_states: torch.Tensor, added_cond_kwargs: Dict[str, Any]
+ ) -> torch.Tensor:
+ if (
+ self.encoder_hid_proj is not None
+ and self.config.encoder_hid_dim_type == "text_proj"
+ ):
+ encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states)
+ elif (
+ self.encoder_hid_proj is not None
+ and self.config.encoder_hid_dim_type == "text_image_proj"
+ ):
+ # Kadinsky 2.1 - style
+ if "image_embeds" not in added_cond_kwargs:
+ raise ValueError(
+ f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`"
+ )
+
+ image_embeds = added_cond_kwargs.get("image_embeds")
+ encoder_hidden_states = self.encoder_hid_proj(
+ encoder_hidden_states, image_embeds
+ )
+ elif (
+ self.encoder_hid_proj is not None
+ and self.config.encoder_hid_dim_type == "image_proj"
+ ):
+ # Kandinsky 2.2 - style
+ if "image_embeds" not in added_cond_kwargs:
+ raise ValueError(
+ f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`"
+ )
+ image_embeds = added_cond_kwargs.get("image_embeds")
+ encoder_hidden_states = self.encoder_hid_proj(image_embeds)
+ elif (
+ self.encoder_hid_proj is not None
+ and self.config.encoder_hid_dim_type == "ip_image_proj"
+ ):
+ if "image_embeds" not in added_cond_kwargs:
+ raise ValueError(
+ f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'ip_image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`"
+ )
+ image_embeds = added_cond_kwargs.get("image_embeds")
+ image_embeds = self.encoder_hid_proj(image_embeds)
+ encoder_hidden_states = (encoder_hidden_states, image_embeds)
+ return encoder_hidden_states
+
+ def forward(
+ self,
+ sample: torch.FloatTensor,
+ timestep: Union[torch.Tensor, float, int],
+ encoder_hidden_states: torch.Tensor,
+ class_labels: Optional[torch.Tensor] = None,
+ timestep_cond: Optional[torch.Tensor] = None,
+ attention_mask: Optional[torch.Tensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+ down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+ mid_block_additional_residual: Optional[torch.Tensor] = None,
+ down_intrablock_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+ encoder_attention_mask: Optional[torch.Tensor] = None,
+ return_dict: bool = True,
+ down_block_add_samples: Optional[Tuple[torch.Tensor]] = None,
+ mid_block_add_sample: Optional[Tuple[torch.Tensor]] = None,
+ up_block_add_samples: Optional[Tuple[torch.Tensor]] = None,
+ ) -> Union[UNet2DConditionOutput, Tuple]:
+ r"""
+ The [`UNet2DConditionModel`] forward method.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The noisy input tensor with the following shape `(batch, channel, height, width)`.
+ timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
+ encoder_hidden_states (`torch.FloatTensor`):
+ The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
+ class_labels (`torch.Tensor`, *optional*, defaults to `None`):
+ Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
+ timestep_cond: (`torch.Tensor`, *optional*, defaults to `None`):
+ Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed
+ through the `self.time_embedding` layer to obtain the timestep embeddings.
+ attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
+ An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+ is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+ negative values to the attention scores corresponding to "discard" tokens.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+ `self.processor` in
+ [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ added_cond_kwargs: (`dict`, *optional*):
+ A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that
+ are passed along to the UNet blocks.
+ down_block_additional_residuals: (`tuple` of `torch.Tensor`, *optional*):
+ A tuple of tensors that if specified are added to the residuals of down unet blocks.
+ mid_block_additional_residual: (`torch.Tensor`, *optional*):
+ A tensor that if specified is added to the residual of the middle unet block.
+ encoder_attention_mask (`torch.Tensor`):
+ A cross-attention mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If
+ `True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias,
+ which adds large negative values to the attention scores corresponding to "discard" tokens.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+ tuple.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttnProcessor`].
+ added_cond_kwargs: (`dict`, *optional*):
+ A kwargs dictionary containin additional embeddings that if specified are added to the embeddings that
+ are passed along to the UNet blocks.
+ down_block_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
+ additional residuals to be added to UNet long skip connections from down blocks to up blocks for
+ example from ControlNet side model(s)
+ mid_block_additional_residual (`torch.Tensor`, *optional*):
+ additional residual to be added to UNet mid block output, for example from ControlNet side model
+ down_intrablock_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
+ additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s)
+
+ Returns:
+ [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
+ If `return_dict` is True, an [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] is returned, otherwise
+ a `tuple` is returned where the first element is the sample tensor.
+ """
+ # By default samples have to be AT least a multiple of the overall upsampling factor.
+ # The overall upsampling factor is equal to 2 ** (# num of upsampling layers).
+ # However, the upsampling interpolation output size can be forced to fit any upsampling size
+ # on the fly if necessary.
+ default_overall_up_factor = 2**self.num_upsamplers
+
+ # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+ forward_upsample_size = False
+ upsample_size = None
+
+ for dim in sample.shape[-2:]:
+ if dim % default_overall_up_factor != 0:
+ # Forward upsample size to force interpolation output size.
+ forward_upsample_size = True
+ break
+
+ # ensure attention_mask is a bias, and give it a singleton query_tokens dimension
+ # expects mask of shape:
+ # [batch, key_tokens]
+ # adds singleton query_tokens dimension:
+ # [batch, 1, key_tokens]
+ # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+ # [batch, heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+ # [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+ if attention_mask is not None:
+ # assume that mask is expressed as:
+ # (1 = keep, 0 = discard)
+ # convert mask into a bias that can be added to attention scores:
+ # (keep = +0, discard = -10000.0)
+ attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+ attention_mask = attention_mask.unsqueeze(1)
+
+ # convert encoder_attention_mask to a bias the same way we do for attention_mask
+ if encoder_attention_mask is not None:
+ encoder_attention_mask = (
+ 1 - encoder_attention_mask.to(sample.dtype)
+ ) * -10000.0
+ encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+ # 0. center input if necessary
+ if self.config.center_input_sample:
+ sample = 2 * sample - 1.0
+
+ # 1. time
+ t_emb = self.get_time_embed(sample=sample, timestep=timestep)
+ emb = self.time_embedding(t_emb, timestep_cond)
+ aug_emb = None
+
+ class_emb = self.get_class_embed(sample=sample, class_labels=class_labels)
+ if class_emb is not None:
+ if self.config.class_embeddings_concat:
+ emb = torch.cat([emb, class_emb], dim=-1)
+ else:
+ emb = emb + class_emb
+
+ aug_emb = self.get_aug_embed(
+ emb=emb,
+ encoder_hidden_states=encoder_hidden_states,
+ added_cond_kwargs=added_cond_kwargs,
+ )
+ if self.config.addition_embed_type == "image_hint":
+ aug_emb, hint = aug_emb
+ sample = torch.cat([sample, hint], dim=1)
+
+ emb = emb + aug_emb if aug_emb is not None else emb
+
+ if self.time_embed_act is not None:
+ emb = self.time_embed_act(emb)
+
+ encoder_hidden_states = self.process_encoder_hidden_states(
+ encoder_hidden_states=encoder_hidden_states,
+ added_cond_kwargs=added_cond_kwargs,
+ )
+
+ # 2. pre-process
+ sample = self.conv_in(sample)
+
+ # 2.5 GLIGEN position net
+ if (
+ cross_attention_kwargs is not None
+ and cross_attention_kwargs.get("gligen", None) is not None
+ ):
+ cross_attention_kwargs = cross_attention_kwargs.copy()
+ gligen_args = cross_attention_kwargs.pop("gligen")
+ cross_attention_kwargs["gligen"] = {
+ "objs": self.position_net(**gligen_args)
+ }
+
+ # 3. down
+ lora_scale = (
+ cross_attention_kwargs.get("scale", 1.0)
+ if cross_attention_kwargs is not None
+ else 1.0
+ )
+ if USE_PEFT_BACKEND:
+ # weight the lora layers by setting `lora_scale` for each PEFT layer
+ scale_lora_layers(self, lora_scale)
+
+ is_controlnet = (
+ mid_block_additional_residual is not None
+ and down_block_additional_residuals is not None
+ )
+ # using new arg down_intrablock_additional_residuals for T2I-Adapters, to distinguish from controlnets
+ is_adapter = down_intrablock_additional_residuals is not None
+ # maintain backward compatibility for legacy usage, where
+ # T2I-Adapter and ControlNet both use down_block_additional_residuals arg
+ # but can only use one or the other
+ is_brushnet = (
+ down_block_add_samples is not None
+ and mid_block_add_sample is not None
+ and up_block_add_samples is not None
+ )
+ if (
+ not is_adapter
+ and mid_block_additional_residual is None
+ and down_block_additional_residuals is not None
+ ):
+ deprecate(
+ "T2I should not use down_block_additional_residuals",
+ "1.3.0",
+ "Passing intrablock residual connections with `down_block_additional_residuals` is deprecated \
+ and will be removed in diffusers 1.3.0. `down_block_additional_residuals` should only be used \
+ for ControlNet. Please make sure use `down_intrablock_additional_residuals` instead. ",
+ standard_warn=False,
+ )
+ down_intrablock_additional_residuals = down_block_additional_residuals
+ is_adapter = True
+
+ down_block_res_samples = (sample,)
+
+ if is_brushnet:
+ sample = sample + down_block_add_samples.pop(0)
+
+ for downsample_block in self.down_blocks:
+ if (
+ hasattr(downsample_block, "has_cross_attention")
+ and downsample_block.has_cross_attention
+ ):
+ # For t2i-adapter CrossAttnDownBlock2D
+ additional_residuals = {}
+ if is_adapter and len(down_intrablock_additional_residuals) > 0:
+ additional_residuals["additional_residuals"] = (
+ down_intrablock_additional_residuals.pop(0)
+ )
+
+ if is_brushnet and len(down_block_add_samples) > 0:
+ additional_residuals["down_block_add_samples"] = [
+ down_block_add_samples.pop(0)
+ for _ in range(
+ len(downsample_block.resnets)
+ + (downsample_block.downsamplers != None)
+ )
+ ]
+
+ sample, res_samples = downsample_block(
+ hidden_states=sample,
+ temb=emb,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=attention_mask,
+ cross_attention_kwargs=cross_attention_kwargs,
+ encoder_attention_mask=encoder_attention_mask,
+ **additional_residuals,
+ )
+ else:
+ additional_residuals = {}
+ if is_brushnet and len(down_block_add_samples) > 0:
+ additional_residuals["down_block_add_samples"] = [
+ down_block_add_samples.pop(0)
+ for _ in range(
+ len(downsample_block.resnets)
+ + (downsample_block.downsamplers != None)
+ )
+ ]
+
+ sample, res_samples = downsample_block(
+ hidden_states=sample,
+ temb=emb,
+ scale=lora_scale,
+ **additional_residuals,
+ )
+ if is_adapter and len(down_intrablock_additional_residuals) > 0:
+ sample += down_intrablock_additional_residuals.pop(0)
+
+ down_block_res_samples += res_samples
+
+ if is_controlnet:
+ new_down_block_res_samples = ()
+
+ for down_block_res_sample, down_block_additional_residual in zip(
+ down_block_res_samples, down_block_additional_residuals
+ ):
+ down_block_res_sample = (
+ down_block_res_sample + down_block_additional_residual
+ )
+ new_down_block_res_samples = new_down_block_res_samples + (
+ down_block_res_sample,
+ )
+
+ down_block_res_samples = new_down_block_res_samples
+
+ # 4. mid
+ if self.mid_block is not None:
+ if (
+ hasattr(self.mid_block, "has_cross_attention")
+ and self.mid_block.has_cross_attention
+ ):
+ sample = self.mid_block(
+ sample,
+ emb,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=attention_mask,
+ cross_attention_kwargs=cross_attention_kwargs,
+ encoder_attention_mask=encoder_attention_mask,
+ )
+ else:
+ sample = self.mid_block(sample, emb)
+
+ # To support T2I-Adapter-XL
+ if (
+ is_adapter
+ and len(down_intrablock_additional_residuals) > 0
+ and sample.shape == down_intrablock_additional_residuals[0].shape
+ ):
+ sample += down_intrablock_additional_residuals.pop(0)
+
+ if is_controlnet:
+ sample = sample + mid_block_additional_residual
+
+ if is_brushnet:
+ sample = sample + mid_block_add_sample
+
+ # 5. up
+ for i, upsample_block in enumerate(self.up_blocks):
+ is_final_block = i == len(self.up_blocks) - 1
+
+ res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+ down_block_res_samples = down_block_res_samples[
+ : -len(upsample_block.resnets)
+ ]
+
+ # if we have not reached the final block and need to forward the
+ # upsample size, we do it here
+ if not is_final_block and forward_upsample_size:
+ upsample_size = down_block_res_samples[-1].shape[2:]
+
+ if (
+ hasattr(upsample_block, "has_cross_attention")
+ and upsample_block.has_cross_attention
+ ):
+ additional_residuals = {}
+ if is_brushnet and len(up_block_add_samples) > 0:
+ additional_residuals["up_block_add_samples"] = [
+ up_block_add_samples.pop(0)
+ for _ in range(
+ len(upsample_block.resnets)
+ + (upsample_block.upsamplers != None)
+ )
+ ]
+
+ sample = upsample_block(
+ hidden_states=sample,
+ temb=emb,
+ res_hidden_states_tuple=res_samples,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ upsample_size=upsample_size,
+ attention_mask=attention_mask,
+ encoder_attention_mask=encoder_attention_mask,
+ **additional_residuals,
+ )
+ else:
+ additional_residuals = {}
+ if is_brushnet and len(up_block_add_samples) > 0:
+ additional_residuals["up_block_add_samples"] = [
+ up_block_add_samples.pop(0)
+ for _ in range(
+ len(upsample_block.resnets)
+ + (upsample_block.upsamplers != None)
+ )
+ ]
+
+ sample = upsample_block(
+ hidden_states=sample,
+ temb=emb,
+ res_hidden_states_tuple=res_samples,
+ upsample_size=upsample_size,
+ scale=lora_scale,
+ **additional_residuals,
+ )
+
+ # 6. post-process
+ if self.conv_norm_out:
+ sample = self.conv_norm_out(sample)
+ sample = self.conv_act(sample)
+ sample = self.conv_out(sample)
+
+ if USE_PEFT_BACKEND:
+ # remove `lora_scale` from each PEFT layer
+ unscale_lora_layers(self, lora_scale)
+
+ if not return_dict:
+ return (sample,)
+
+ return UNet2DConditionOutput(sample=sample)
diff --git a/iopaint/model_manager.py b/iopaint/model_manager.py
index fdd2c46..d66b8fe 100644
--- a/iopaint/model_manager.py
+++ b/iopaint/model_manager.py
@@ -8,6 +8,7 @@ from iopaint.download import scan_models
from iopaint.helper import switch_mps_device
from iopaint.model import models, ControlNet, SD, SDXL
from iopaint.model.brushnet.brushnet_wrapper import BrushNetWrapper
+from iopaint.model.power_paint.power_paint_v2 import PowerPaintV2
from iopaint.model.utils import torch_gc, is_local_files_only
from iopaint.schema import InpaintRequest, ModelInfo, ModelType
@@ -23,9 +24,9 @@ class ModelManager:
self.enable_controlnet = kwargs.get("enable_controlnet", False)
controlnet_method = kwargs.get("controlnet_method", None)
if (
- controlnet_method is None
- and name in self.available_models
- and self.available_models[name].support_controlnet
+ controlnet_method is None
+ and name in self.available_models
+ and self.available_models[name].support_controlnet
):
controlnet_method = self.available_models[name].controlnets[0]
self.controlnet_method = controlnet_method
@@ -33,6 +34,8 @@ class ModelManager:
self.enable_brushnet = kwargs.get("enable_brushnet", False)
self.brushnet_method = kwargs.get("brushnet_method", None)
+ self.enable_powerpaint_v2 = kwargs.get("enable_powerpaint_v2", False)
+
self.model = self.init_model(name, device, **kwargs)
@property
@@ -62,6 +65,9 @@ class ModelManager:
if model_info.support_brushnet and self.enable_brushnet:
return BrushNetWrapper(device, **kwargs)
+ if model_info.support_powerpaint_v2 and self.enable_powerpaint_v2:
+ return PowerPaintV2(device, **kwargs)
+
if model_info.name in models:
return models[name](device, **kwargs)
@@ -91,10 +97,12 @@ class ModelManager:
Returns:
BGR image
"""
- if not config.enable_brushnet:
+ if config.enable_controlnet:
self.switch_controlnet_method(config)
- if not config.enable_controlnet:
+ if config.enable_brushnet:
self.switch_brushnet_method(config)
+
+ self.enable_disable_powerpaint_v2(config)
self.enable_disable_freeu(config)
self.enable_disable_lcm_lora(config)
return self.model(image, mask, config).astype(np.uint8)
@@ -113,9 +121,9 @@ class ModelManager:
self.name = new_name
if (
- self.available_models[new_name].support_controlnet
- and self.controlnet_method
- not in self.available_models[new_name].controlnets
+ self.available_models[new_name].support_controlnet
+ and self.controlnet_method
+ not in self.available_models[new_name].controlnets
):
self.controlnet_method = self.available_models[new_name].controlnets[0]
try:
@@ -140,9 +148,9 @@ class ModelManager:
return
if (
- self.enable_brushnet
- and config.brushnet_method
- and self.brushnet_method != config.brushnet_method
+ self.enable_brushnet
+ and config.brushnet_method
+ and self.brushnet_method != config.brushnet_method
):
old_brushnet_method = self.brushnet_method
self.brushnet_method = config.brushnet_method
@@ -180,9 +188,9 @@ class ModelManager:
return
if (
- self.enable_controlnet
- and config.controlnet_method
- and self.controlnet_method != config.controlnet_method
+ self.enable_controlnet
+ and config.controlnet_method
+ and self.controlnet_method != config.controlnet_method
):
old_controlnet_method = self.controlnet_method
self.controlnet_method = config.controlnet_method
@@ -213,6 +221,25 @@ class ModelManager:
else:
logger.info(f"Enable controlnet: {config.controlnet_method}")
+ def enable_disable_powerpaint_v2(self, config: InpaintRequest):
+ if not self.available_models[self.name].support_powerpaint_v2:
+ return
+
+ if self.enable_powerpaint_v2 != config.enable_powerpaint_v2:
+ self.enable_powerpaint_v2 = config.enable_powerpaint_v2
+ pipe_components = {"vae": self.model.model.vae}
+
+ self.model = self.init_model(
+ self.name,
+ switch_mps_device(self.name, self.device),
+ pipe_components=pipe_components,
+ **self.kwargs,
+ )
+ if config.enable_powerpaint_v2:
+ logger.info("Enable PowerPaintV2")
+ else:
+ logger.info("Disable PowerPaintV2")
+
def enable_disable_freeu(self, config: InpaintRequest):
if str(self.model.device) == "mps":
return
diff --git a/iopaint/schema.py b/iopaint/schema.py
index 24c115b..1b87f14 100644
--- a/iopaint/schema.py
+++ b/iopaint/schema.py
@@ -119,6 +119,13 @@ class ModelInfo(BaseModel):
ModelType.DIFFUSERS_SD,
]
+ @computed_field
+ @property
+ def support_powerpaint_v2(self) -> bool:
+ return self.model_type in [
+ ModelType.DIFFUSERS_SD,
+ ]
+
@computed_field
@property
def support_freeu(self) -> bool:
@@ -225,8 +232,9 @@ class FREEUConfig(BaseModel):
b2: float = 1.4
-class PowerPaintTask(str, Enum):
+class PowerPaintTask(Choices):
text_guided = "text-guided"
+ context_aware = "context-aware"
shape_guided = "shape-guided"
object_remove = "object-remove"
outpainting = "outpainting"
@@ -387,12 +395,13 @@ class InpaintRequest(BaseModel):
# BrushNet
enable_brushnet: bool = Field(False, description="Enable brushnet")
- brushnet_method: str = Field(
- SD_BRUSHNET_CHOICES[0], description="Brushnet method"
+ brushnet_method: str = Field(SD_BRUSHNET_CHOICES[0], description="Brushnet method")
+ brushnet_conditioning_scale: float = Field(
+ 1.0, description="brushnet conditioning scale", ge=0.0, le=1.0
)
- brushnet_conditioning_scale: float = Field(1.0, description="brushnet conditioning scale", ge=0.0, le=1.0)
# PowerPaint
+ enable_powerpaint_v2: bool = Field(False, description="Enable PowerPaint v2")
powerpaint_task: PowerPaintTask = Field(
PowerPaintTask.text_guided, description="PowerPaint task"
)
@@ -403,8 +412,8 @@ class InpaintRequest(BaseModel):
le=1.0,
)
- @model_validator(mode='after')
- def validate_field(cls, values: 'InpaintRequest'):
+ @model_validator(mode="after")
+ def validate_field(cls, values: "InpaintRequest"):
if values.sd_seed == -1:
values.sd_seed = random.randint(1, 99999999)
logger.info(f"Generate random seed: {values.sd_seed}")
diff --git a/iopaint/tests/test_brushnet.py b/iopaint/tests/test_brushnet.py
index 0261b1d..8da4aa1 100644
--- a/iopaint/tests/test_brushnet.py
+++ b/iopaint/tests/test_brushnet.py
@@ -10,7 +10,7 @@ import pytest
import torch
from iopaint.model_manager import ModelManager
-from iopaint.schema import HDStrategy, SDSampler, FREEUConfig
+from iopaint.schema import HDStrategy, SDSampler, FREEUConfig, PowerPaintTask
current_dir = Path(__file__).parent.absolute().resolve()
save_dir = current_dir / "result"
@@ -35,7 +35,7 @@ def test_runway_brushnet(device, sampler):
sd_freeu=True,
sd_freeu_config=FREEUConfig(),
enable_brushnet=True,
- brushnet_method=SD_BRUSHNET_CHOICES[0]
+ brushnet_method=SD_BRUSHNET_CHOICES[0],
)
cfg.sd_sampler = sampler
@@ -49,38 +49,64 @@ def test_runway_brushnet(device, sampler):
@pytest.mark.parametrize("device", ["cuda", "mps"])
-@pytest.mark.parametrize("sampler", [SDSampler.dpm_plus_plus_2m_karras])
-@pytest.mark.parametrize(
- "name",
- [
- "v1-5-pruned-emaonly.safetensors",
- ],
-)
-def test_brushnet_local_file_path(device, sampler, name):
+@pytest.mark.parametrize("sampler", [SDSampler.dpm_plus_plus_2m])
+def test_runway_powerpaint_v2(device, sampler):
sd_steps = check_device(device)
model = ModelManager(
- name=name,
+ name="runwayml/stable-diffusion-v1-5",
device=torch.device(device),
disable_nsfw=True,
sd_cpu_textencoder=False,
- cpu_offload=False,
)
- cfg = get_config(
- strategy=HDStrategy.ORIGINAL,
- prompt="face of a fox, sitting on a bench",
- sd_steps=sd_steps,
- sd_seed=1234,
- enable_brushnet=True,
- brushnet_method=SD_BRUSHNET_CHOICES[1]
- )
- cfg.sd_sampler = sampler
- assert_equal(
- model,
- cfg,
- f"brushnet_segmentation_mask_{device}.png",
- img_p=current_dir / "overture-creations-5sI6fQgYIuo.png",
- mask_p=current_dir / "overture-creations-5sI6fQgYIuo_mask.png",
- fx=1,
- fy=1,
- )
+ tasks = {
+ PowerPaintTask.text_guided: {
+ "prompt": "face of a fox, sitting on a bench",
+ "scale": 7.5,
+ },
+ PowerPaintTask.context_aware: {
+ "prompt": "face of a fox, sitting on a bench",
+ "scale": 7.5,
+ },
+ PowerPaintTask.shape_guided: {
+ "prompt": "face of a fox, sitting on a bench",
+ "scale": 7.5,
+ },
+ PowerPaintTask.object_remove: {
+ "prompt": "",
+ "scale": 12,
+ },
+ PowerPaintTask.outpainting: {
+ "prompt": "",
+ "scale": 7.5,
+ },
+ }
+
+ for task, data in tasks.items():
+ cfg = get_config(
+ strategy=HDStrategy.ORIGINAL,
+ prompt=data["prompt"],
+ negative_prompt="out of frame, lowres, error, cropped, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, out of frame, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, disfigured, gross proportions, malformed limbs, watermark, signature",
+ sd_steps=sd_steps,
+ sd_guidance_scale=data["scale"],
+ enable_powerpaint_v2=True,
+ powerpaint_task=task,
+ sd_sampler=sampler,
+ sd_mask_blur=11,
+ sd_seed=42,
+ # sd_keep_unmasked_area=False
+ )
+ if task == PowerPaintTask.outpainting:
+ cfg.use_extender = True
+ cfg.extender_x = -128
+ cfg.extender_y = -128
+ cfg.extender_width = 768
+ cfg.extender_height = 768
+
+ assert_equal(
+ model,
+ cfg,
+ f"powerpaint_v2_{device}_{task}.png",
+ img_p=current_dir / "overture-creations-5sI6fQgYIuo.png",
+ mask_p=current_dir / "overture-creations-5sI6fQgYIuo_mask.png",
+ )
diff --git a/iopaint/tests/test_outpainting.py b/iopaint/tests/test_outpainting.py
index 024d701..ce48751 100644
--- a/iopaint/tests/test_outpainting.py
+++ b/iopaint/tests/test_outpainting.py
@@ -3,18 +3,17 @@ import os
from iopaint.tests.utils import current_dir, check_device
os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "1"
-from pathlib import Path
import pytest
import torch
from iopaint.model_manager import ModelManager
-from iopaint.schema import HDStrategy, SDSampler
+from iopaint.schema import SDSampler
from iopaint.tests.test_model import get_config, assert_equal
@pytest.mark.parametrize("name", ["runwayml/stable-diffusion-inpainting"])
-@pytest.mark.parametrize("device", ["cuda", "mps", "cpu"])
+@pytest.mark.parametrize("device", ["cuda", "mps"])
@pytest.mark.parametrize(
"rect",
[
@@ -23,7 +22,7 @@ from iopaint.tests.test_model import get_config, assert_equal
[128, 0, 512 - 128 + 100, 512],
[-100, 0, 512 - 128 + 100, 512],
[0, 0, 512, 512 + 200],
- [0, 0, 512 + 200, 512],
+ [256, 0, 512 + 200, 512],
[-100, -100, 512 + 200, 512 + 200],
],
)
@@ -58,7 +57,7 @@ def test_outpainting(name, device, rect):
@pytest.mark.parametrize("name", ["kandinsky-community/kandinsky-2-2-decoder-inpaint"])
-@pytest.mark.parametrize("device", ["cuda", "mps", "cpu"])
+@pytest.mark.parametrize("device", ["cuda", "mps"])
@pytest.mark.parametrize(
"rect",
[
@@ -99,7 +98,7 @@ def test_kandinsky_outpainting(name, device, rect):
@pytest.mark.parametrize("name", ["Sanster/PowerPaint-V1-stable-diffusion-inpainting"])
-@pytest.mark.parametrize("device", ["cuda", "mps", "cpu"])
+@pytest.mark.parametrize("device", ["cuda", "mps"])
@pytest.mark.parametrize(
"rect",
[
@@ -114,7 +113,7 @@ def test_powerpaint_outpainting(name, device, rect):
device=torch.device(device),
disable_nsfw=True,
sd_cpu_textencoder=False,
- low_mem=True
+ low_mem=True,
)
cfg = get_config(
prompt="a dog sitting on a bench in the park",
diff --git a/iopaint/tests/utils.py b/iopaint/tests/utils.py
index 08f4aeb..92bc5ef 100644
--- a/iopaint/tests/utils.py
+++ b/iopaint/tests/utils.py
@@ -3,9 +3,8 @@ import cv2
import pytest
import torch
-from iopaint.helper import encode_pil_to_base64
from iopaint.schema import LDMSampler, HDStrategy, InpaintRequest, SDSampler
-from PIL import Image
+import numpy as np
current_dir = Path(__file__).parent.absolute().resolve()
save_dir = current_dir / "result"
@@ -32,6 +31,7 @@ def assert_equal(
):
img, mask = get_data(fx=fx, fy=fy, img_p=img_p, mask_p=mask_p)
print(f"Input image shape: {img.shape}")
+
res = model(img, mask, config)
ok = cv2.imwrite(
str(save_dir / gt_name),