IOPaint/iopaint/model/sdxl.py

111 lines
3.9 KiB
Python
Raw Normal View History

2023-12-01 03:15:35 +01:00
import os
2023-11-14 07:19:56 +01:00
import PIL.Image
import cv2
import torch
2023-12-01 03:15:35 +01:00
from diffusers import AutoencoderKL
2023-11-14 07:19:56 +01:00
from loguru import logger
2024-01-05 08:19:23 +01:00
from iopaint.schema import InpaintRequest, ModelType
2023-11-14 07:19:56 +01:00
2024-01-05 09:40:06 +01:00
from .base import DiffusionInpaintModel
2024-01-10 06:34:11 +01:00
from .helper.cpu_text_encoder import CPUTextEncoderWrapper
from .original_sd_configs import get_config_files
from .utils import (
handle_from_pretrained_exceptions,
get_torch_dtype,
enable_low_mem,
is_local_files_only,
)
2024-01-05 09:40:06 +01:00
2023-11-14 07:19:56 +01:00
class SDXL(DiffusionInpaintModel):
2023-12-27 15:00:07 +01:00
name = "diffusers/stable-diffusion-xl-1.0-inpainting-0.1"
2023-11-14 07:19:56 +01:00
pad_mod = 8
min_size = 512
2023-11-15 01:50:35 +01:00
lcm_lora_id = "latent-consistency/lcm-lora-sdxl"
2023-12-01 03:15:35 +01:00
model_id_or_path = "diffusers/stable-diffusion-xl-1.0-inpainting-0.1"
2023-11-14 07:19:56 +01:00
def init_model(self, device: torch.device, **kwargs):
2023-12-01 03:15:35 +01:00
from diffusers.pipelines import StableDiffusionXLInpaintPipeline
2023-11-14 07:19:56 +01:00
use_gpu, torch_dtype = get_torch_dtype(device, kwargs.get("no_half", False))
2023-11-14 07:19:56 +01:00
2023-12-15 05:40:29 +01:00
if self.model_info.model_type == ModelType.DIFFUSERS_SDXL:
num_in_channels = 4
else:
num_in_channels = 9
2023-12-01 03:15:35 +01:00
if os.path.isfile(self.model_id_or_path):
self.model = StableDiffusionXLInpaintPipeline.from_single_file(
2023-12-15 05:40:29 +01:00
self.model_id_or_path,
2023-12-27 15:00:07 +01:00
dtype=torch_dtype,
2023-12-15 05:40:29 +01:00
num_in_channels=num_in_channels,
load_safety_checker=False,
config_files=get_config_files()
2023-12-01 03:15:35 +01:00
)
else:
model_kwargs = {
**kwargs.get("pipe_components", {}),
"local_files_only": is_local_files_only(**kwargs),
}
2024-01-10 06:34:11 +01:00
if "vae" not in model_kwargs:
vae = AutoencoderKL.from_pretrained(
"madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch_dtype
)
2024-01-10 06:34:11 +01:00
model_kwargs["vae"] = vae
2023-12-27 15:00:07 +01:00
self.model = handle_from_pretrained_exceptions(
StableDiffusionXLInpaintPipeline.from_pretrained,
pretrained_model_name_or_path=self.model_id_or_path,
2023-12-01 03:15:35 +01:00
torch_dtype=torch_dtype,
2023-12-27 15:00:07 +01:00
variant="fp16",
**model_kwargs
2023-12-01 03:15:35 +01:00
)
2023-11-14 07:19:56 +01:00
2024-01-09 15:42:48 +01:00
enable_low_mem(self.model, kwargs.get("low_mem", False))
2023-11-14 07:19:56 +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"]:
2024-01-10 06:34:11 +01:00
logger.info("Run Stable Diffusion TextEncoder on CPU")
self.model.text_encoder = CPUTextEncoderWrapper(
self.model.text_encoder, torch_dtype
)
self.model.text_encoder_2 = CPUTextEncoderWrapper(
self.model.text_encoder_2, torch_dtype
)
2023-11-14 07:19:56 +01:00
self.callback = kwargs.pop("callback", None)
2023-12-30 16:36:44 +01:00
def forward(self, image, mask, config: InpaintRequest):
2023-11-14 07:19:56 +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
"""
2023-11-15 01:50:35 +01:00
self.set_scheduler(config)
2023-11-14 07:19:56 +01:00
img_h, img_w = image.shape[:2]
output = self.model(
image=PIL.Image.fromarray(image),
prompt=config.prompt,
negative_prompt=config.negative_prompt,
mask_image=PIL.Image.fromarray(mask[:, :, -1], mode="L"),
num_inference_steps=config.sd_steps,
strength=0.999 if config.sd_strength == 1.0 else config.sd_strength,
guidance_scale=config.sd_guidance_scale,
output_type="np",
2024-01-02 10:13:11 +01:00
callback_on_step_end=self.callback,
2023-11-14 07:19:56 +01:00
height=img_h,
width=img_w,
generator=torch.manual_seed(config.sd_seed),
).images[0]
output = (output * 255).round().astype("uint8")
output = cv2.cvtColor(output, cv2.COLOR_RGB2BGR)
return output