IOPaint/iopaint/model/controlnet.py

187 lines
6.6 KiB
Python
Raw Normal View History

2023-03-19 15:40:23 +01:00
import PIL.Image
import cv2
import torch
2024-01-05 09:40:06 +01:00
from diffusers import ControlNetModel
2023-03-19 15:40:23 +01:00
from loguru import logger
2024-01-05 09:40:06 +01:00
from iopaint.schema import InpaintRequest, ModelType
2023-03-19 15:40:23 +01:00
2024-01-05 09:40:06 +01:00
from .base import DiffusionInpaintModel
from .helper.controlnet_preprocess import (
2023-12-01 03:15:35 +01:00
make_canny_control_image,
make_openpose_control_image,
make_depth_control_image,
make_inpaint_control_image,
)
2024-01-05 09:40:06 +01:00
from .helper.cpu_text_encoder import CPUTextEncoderWrapper
from .original_sd_configs import get_config_files
2024-01-09 15:42:48 +01:00
from .utils import (
get_scheduler,
handle_from_pretrained_exceptions,
get_torch_dtype,
enable_low_mem,
is_local_files_only,
2024-01-09 15:42:48 +01:00
)
2023-12-01 03:15:35 +01:00
2023-04-01 03:23:37 +02:00
2023-03-19 15:40:23 +01:00
class ControlNet(DiffusionInpaintModel):
name = "controlnet"
pad_mod = 8
min_size = 512
2023-12-19 06:16:30 +01:00
@property
def lcm_lora_id(self):
if self.model_info.model_type in [
ModelType.DIFFUSERS_SD,
ModelType.DIFFUSERS_SD_INPAINT,
]:
return "latent-consistency/lcm-lora-sdv1-5"
if self.model_info.model_type in [
ModelType.DIFFUSERS_SDXL,
ModelType.DIFFUSERS_SDXL_INPAINT,
]:
return "latent-consistency/lcm-lora-sdxl"
raise NotImplementedError(f"Unsupported controlnet lcm model {self.model_info}")
2023-03-19 15:40:23 +01:00
def init_model(self, device: torch.device, **kwargs):
2024-01-02 10:13:11 +01:00
model_info = kwargs["model_info"]
2023-12-27 15:00:07 +01:00
controlnet_method = kwargs["controlnet_method"]
2023-12-01 03:15:35 +01:00
self.model_info = model_info
2023-12-27 15:00:07 +01:00
self.controlnet_method = controlnet_method
2023-03-19 15:40:23 +01:00
model_kwargs = {
**kwargs.get("pipe_components", {}),
"local_files_only": is_local_files_only(**kwargs),
}
self.local_files_only = model_kwargs["local_files_only"]
disable_nsfw_checker = kwargs["disable_nsfw"] or kwargs.get(
"cpu_offload", False
)
if disable_nsfw_checker:
2023-03-19 15:40:23 +01:00
logger.info("Disable Stable Diffusion Model NSFW checker")
model_kwargs.update(
dict(
safety_checker=None,
feature_extractor=None,
requires_safety_checker=False,
)
)
use_gpu, torch_dtype = get_torch_dtype(device, kwargs.get("no_half", False))
2023-12-22 07:00:30 +01:00
self.torch_dtype = torch_dtype
2023-03-19 15:40:23 +01:00
2023-12-01 03:15:35 +01:00
if model_info.model_type in [
ModelType.DIFFUSERS_SD,
ModelType.DIFFUSERS_SD_INPAINT,
]:
from diffusers import (
StableDiffusionControlNetInpaintPipeline as PipeClass,
)
elif model_info.model_type in [
ModelType.DIFFUSERS_SDXL,
ModelType.DIFFUSERS_SDXL_INPAINT,
]:
from diffusers import (
StableDiffusionXLControlNetInpaintPipeline as PipeClass,
)
2023-05-02 08:09:40 +02:00
2023-03-19 15:40:23 +01:00
controlnet = ControlNetModel.from_pretrained(
2023-12-27 15:00:07 +01:00
pretrained_model_name_or_path=controlnet_method,
resume_download=True,
local_files_only=model_kwargs["local_files_only"],
2023-03-19 15:40:23 +01:00
)
2023-12-01 03:15:35 +01:00
if model_info.is_single_file_diffusers:
2023-12-15 05:40:29 +01:00
if self.model_info.model_type == ModelType.DIFFUSERS_SD:
model_kwargs["num_in_channels"] = 4
else:
model_kwargs["num_in_channels"] = 9
2023-12-01 03:15:35 +01:00
self.model = PipeClass.from_single_file(
model_info.path,
controlnet=controlnet,
load_safety_checker=not disable_nsfw_checker,
2024-01-10 14:25:51 +01:00
torch_dtype=torch_dtype,
config_files=get_config_files(),
**model_kwargs,
2024-01-10 14:25:51 +01:00
)
2023-04-01 03:23:37 +02:00
else:
2023-12-27 15:00:07 +01:00
self.model = handle_from_pretrained_exceptions(
PipeClass.from_pretrained,
pretrained_model_name_or_path=model_info.path,
2023-04-01 03:23:37 +02:00
controlnet=controlnet,
2023-12-27 15:00:07 +01:00
variant="fp16",
dtype=torch_dtype,
2023-04-01 03:23:37 +02:00
**model_kwargs,
)
2023-03-19 15:40:23 +01:00
2024-01-08 16:54:20 +01:00
enable_low_mem(self.model, kwargs.get("low_mem", False))
2023-03-19 15:40:23 +01:00
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)
2023-12-22 07:00:30 +01:00
def switch_controlnet_method(self, new_method: str):
2023-12-27 15:00:07 +01:00
self.controlnet_method = new_method
2023-12-22 07:00:30 +01:00
controlnet = ControlNetModel.from_pretrained(
new_method, resume_download=True, local_files_only=self.local_files_only
2023-12-22 07:00:30 +01:00
).to(self.model.device)
self.model.controlnet = controlnet
2023-12-01 03:15:35 +01:00
def _get_control_image(self, image, mask):
2023-12-27 15:00:07 +01:00
if "canny" in self.controlnet_method:
2023-12-01 03:15:35 +01:00
control_image = make_canny_control_image(image)
2023-12-27 15:00:07 +01:00
elif "openpose" in self.controlnet_method:
2023-12-01 03:15:35 +01:00
control_image = make_openpose_control_image(image)
2023-12-27 15:00:07 +01:00
elif "depth" in self.controlnet_method:
2023-12-01 03:15:35 +01:00
control_image = make_depth_control_image(image)
2023-12-27 15:00:07 +01:00
elif "inpaint" in self.controlnet_method:
2023-12-01 03:15:35 +01:00
control_image = make_inpaint_control_image(image, mask)
else:
2023-12-27 15:00:07 +01:00
raise NotImplementedError(f"{self.controlnet_method} not implemented")
2023-12-01 03:15:35 +01:00
return control_image
2023-12-30 16:36:44 +01:00
def forward(self, image, mask, config: InpaintRequest):
2023-03-19 15:40:23 +01:00
"""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
"""
scheduler_config = self.model.scheduler.config
scheduler = get_scheduler(config.sd_sampler, scheduler_config)
self.model.scheduler = scheduler
img_h, img_w = image.shape[:2]
2023-12-01 03:15:35 +01:00
control_image = self._get_control_image(image, mask)
mask_image = PIL.Image.fromarray(mask[:, :, -1], mode="L")
image = PIL.Image.fromarray(image)
output = self.model(
image=image,
mask_image=mask_image,
control_image=control_image,
prompt=config.prompt,
negative_prompt=config.negative_prompt,
num_inference_steps=config.sd_steps,
guidance_scale=config.sd_guidance_scale,
output_type="np",
2024-01-02 10:13:11 +01:00
callback_on_step_end=self.callback,
2023-12-01 03:15:35 +01:00
height=img_h,
width=img_w,
generator=torch.manual_seed(config.sd_seed),
controlnet_conditioning_scale=config.controlnet_conditioning_scale,
).images[0]
2023-03-19 15:40:23 +01:00
output = (output * 255).round().astype("uint8")
output = cv2.cvtColor(output, cv2.COLOR_RGB2BGR)
return output