IOPaint/lama_cleaner/model/sd.py

147 lines
5.0 KiB
Python
Raw Normal View History

2023-12-01 03:15:35 +01:00
import os
2023-03-29 16:05:34 +02:00
2022-09-15 16:21:27 +02:00
import PIL.Image
import cv2
import numpy as np
import torch
from loguru import logger
2023-12-01 03:15:35 +01:00
from lama_cleaner.const import DIFFUSERS_MODEL_FP16_REVERSION
2023-01-27 13:59:22 +01:00
from lama_cleaner.model.base import DiffusionInpaintModel
2023-12-01 03:15:35 +01:00
from lama_cleaner.model.helper.cpu_text_encoder import CPUTextEncoderWrapper
2023-12-15 05:40:29 +01:00
from lama_cleaner.schema import Config, ModelType
2022-09-15 16:21:27 +02:00
2023-01-27 13:59:22 +01:00
class SD(DiffusionInpaintModel):
pad_mod = 8
2022-09-15 16:21:27 +02:00
min_size = 512
2023-11-15 01:50:35 +01:00
lcm_lora_id = "latent-consistency/lcm-lora-sdv1-5"
2022-09-15 16:21:27 +02:00
def init_model(self, device: torch.device, **kwargs):
2022-10-20 15:01:14 +02:00
from diffusers.pipelines.stable_diffusion import StableDiffusionInpaintPipeline
2022-09-15 16:21:27 +02:00
2023-02-07 14:00:19 +01:00
fp16 = not kwargs.get("no_half", False)
2022-09-29 03:42:19 +02:00
2023-12-01 03:15:35 +01:00
model_kwargs = {}
2023-02-07 14:00:19 +01:00
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,
)
)
use_gpu = device == torch.device("cuda") and torch.cuda.is_available()
2023-01-04 14:27:37 +01:00
torch_dtype = torch.float16 if use_gpu and fp16 else torch.float32
2023-03-29 16:05:34 +02:00
2023-12-15 05:40:29 +01:00
if self.model_info.is_single_file_diffusers:
if self.model_info.model_type == ModelType.DIFFUSERS_SD:
model_kwargs["num_in_channels"] = 4
else:
model_kwargs["num_in_channels"] = 9
2023-11-16 07:09:08 +01:00
self.model = StableDiffusionInpaintPipeline.from_single_file(
2023-12-01 03:15:35 +01:00
self.model_id_or_path, torch_dtype=torch_dtype, **model_kwargs
2023-03-29 16:05:34 +02:00
)
else:
self.model = StableDiffusionInpaintPipeline.from_pretrained(
self.model_id_or_path,
2023-12-01 03:15:35 +01:00
revision="fp16"
if (
self.model_id_or_path in DIFFUSERS_MODEL_FP16_REVERSION
and use_gpu
and fp16
)
else "main",
2023-03-29 16:05:34 +02:00
torch_dtype=torch_dtype,
use_auth_token=kwargs["hf_access_token"],
**model_kwargs,
)
2023-01-05 15:07:39 +01:00
# https://huggingface.co/docs/diffusers/v0.7.0/en/api/pipelines/stable_diffusion#diffusers.StableDiffusionInpaintPipeline.enable_attention_slicing
2022-09-15 16:21:27 +02:00
self.model.enable_attention_slicing()
# https://huggingface.co/docs/diffusers/v0.7.0/en/optimization/fp16#memory-efficient-attention
2023-02-07 14:00:19 +01:00
if kwargs.get("enable_xformers", False):
self.model.enable_xformers_memory_efficient_attention()
2022-09-29 06:20:55 +02:00
2023-02-07 14:00:19 +01:00
if kwargs.get("cpu_offload", False) and use_gpu:
2023-01-05 15:07:39 +01:00
# TODO: gpu_id
2023-01-07 01:52:11 +01:00
logger.info("Enable sequential cpu offload")
2023-01-05 15:07:39 +01:00
self.model.enable_sequential_cpu_offload(gpu_id=0)
else:
self.model = self.model.to(device)
2023-02-07 14:00:19 +01:00
if kwargs["sd_cpu_textencoder"]:
2023-01-05 15:07:39 +01:00
logger.info("Run Stable Diffusion TextEncoder on CPU")
2023-02-07 14:00:19 +01:00
self.model.text_encoder = CPUTextEncoderWrapper(
self.model.text_encoder, torch_dtype
)
2022-09-29 06:20:55 +02:00
2022-10-15 16:32:25 +02:00
self.callback = kwargs.pop("callback", None)
2022-09-15 16:21:27 +02:00
def forward(self, image, mask, config: Config):
"""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
"""
2023-11-15 01:50:35 +01:00
self.set_scheduler(config)
2022-09-22 06:38:32 +02:00
2022-09-22 15:50:41 +02:00
if config.sd_mask_blur != 0:
k = 2 * config.sd_mask_blur + 1
mask = cv2.GaussianBlur(mask, (k, k), 0)[:, :, np.newaxis]
img_h, img_w = image.shape[:2]
2022-09-15 16:21:27 +02:00
output = self.model(
2022-12-04 06:41:48 +01:00
image=PIL.Image.fromarray(image),
2022-09-15 16:21:27 +02:00
prompt=config.prompt,
2022-11-08 14:58:48 +01:00
negative_prompt=config.negative_prompt,
2022-09-15 16:21:27 +02:00
mask_image=PIL.Image.fromarray(mask[:, :, -1], mode="L"),
num_inference_steps=config.sd_steps,
2023-11-14 07:02:15 +01:00
strength=config.sd_strength,
2022-09-15 16:21:27 +02:00
guidance_scale=config.sd_guidance_scale,
2023-08-30 15:30:11 +02:00
output_type="np",
2022-10-15 16:32:25 +02:00
callback=self.callback,
height=img_h,
width=img_w,
2023-03-01 14:44:02 +01:00
generator=torch.manual_seed(config.sd_seed),
2023-11-15 02:10:13 +01:00
callback_steps=1,
2022-09-15 16:21:27 +02:00
).images[0]
output = (output * 255).round().astype("uint8")
output = cv2.cvtColor(output, cv2.COLOR_RGB2BGR)
return output
@staticmethod
def is_downloaded() -> bool:
# model will be downloaded when app start, and can't switch in frontend settings
return True
2023-11-16 14:12:06 +01:00
@classmethod
def download(cls):
from diffusers import StableDiffusionInpaintPipeline
StableDiffusionInpaintPipeline.from_pretrained(cls.model_id_or_path)
2022-09-15 16:21:27 +02:00
class SD15(SD):
2023-02-11 06:30:09 +01:00
name = "sd1.5"
2022-10-20 15:01:14 +02:00
model_id_or_path = "runwayml/stable-diffusion-inpainting"
2022-12-04 06:41:48 +01:00
2023-03-01 14:44:02 +01:00
class Anything4(SD):
name = "anything4"
model_id_or_path = "Sanster/anything-4.0-inpainting"
class RealisticVision14(SD):
name = "realisticVision1.4"
model_id_or_path = "Sanster/Realistic_Vision_V1.4-inpainting"
2022-12-04 06:41:48 +01:00
class SD2(SD):
2023-02-11 06:30:09 +01:00
name = "sd2"
2022-12-04 06:41:48 +01:00
model_id_or_path = "stabilityai/stable-diffusion-2-inpainting"