IOPaint/lama_cleaner/model/sd.py

194 lines
6.5 KiB
Python
Raw Normal View History

2023-03-29 16:05:34 +02:00
import gc
2022-09-15 16:21:27 +02:00
import PIL.Image
import cv2
import numpy as np
import torch
from loguru import logger
2023-01-27 13:59:22 +01:00
from lama_cleaner.model.base import DiffusionInpaintModel
2023-03-19 15:40:23 +01:00
from lama_cleaner.model.utils import torch_gc, get_scheduler
from lama_cleaner.schema import Config
2022-09-15 16:21:27 +02:00
2022-10-20 15:01:14 +02:00
class CPUTextEncoderWrapper:
2022-11-04 08:33:59 +01:00
def __init__(self, text_encoder, torch_dtype):
2022-12-04 15:04:36 +01:00
self.config = text_encoder.config
2023-02-07 14:00:19 +01:00
self.text_encoder = text_encoder.to(torch.device("cpu"), non_blocking=True)
2022-10-20 15:01:14 +02:00
self.text_encoder = self.text_encoder.to(torch.float32, non_blocking=True)
2022-11-04 08:33:59 +01:00
self.torch_dtype = torch_dtype
2023-01-05 15:07:39 +01:00
del text_encoder
torch_gc()
2022-10-20 15:01:14 +02:00
2022-12-04 15:04:36 +01:00
def __call__(self, x, **kwargs):
2022-10-20 15:01:14 +02:00
input_device = x.device
2023-02-07 14:00:19 +01:00
return [
self.text_encoder(x.to(self.text_encoder.device), **kwargs)[0]
.to(input_device)
.to(self.torch_dtype)
]
@property
def dtype(self):
return self.torch_dtype
2022-10-20 15:01:14 +02:00
2023-03-30 15:16:55 +02:00
def load_from_local_model(local_model_path, torch_dtype, disable_nsfw=True):
2023-03-29 16:05:34 +02:00
from diffusers.pipelines.stable_diffusion.convert_from_ckpt import (
load_pipeline_from_original_stable_diffusion_ckpt,
)
from diffusers.pipelines.stable_diffusion import StableDiffusionInpaintPipeline
logger.info(f"Converting {local_model_path} to diffusers pipeline")
pipe = load_pipeline_from_original_stable_diffusion_ckpt(
local_model_path,
num_in_channels=9,
from_safetensors=local_model_path.endswith("safetensors"),
device="cpu",
)
inpaint_pipe = StableDiffusionInpaintPipeline(
vae=pipe.vae,
text_encoder=pipe.text_encoder,
tokenizer=pipe.tokenizer,
unet=pipe.unet,
scheduler=pipe.scheduler,
safety_checker=None if disable_nsfw else pipe.safety_checker,
feature_extractor=None if disable_nsfw else pipe.safety_checker,
requires_safety_checker=not disable_nsfw,
)
del pipe
gc.collect()
return inpaint_pipe.to(torch_dtype)
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
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
2023-02-07 14:00:19 +01:00
model_kwargs = {
"local_files_only": kwargs.get("local_files_only", kwargs["sd_run_local"])
}
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
if kwargs.get("sd_local_model_path", None):
self.model = load_from_local_model(
kwargs["sd_local_model_path"],
torch_dtype=torch_dtype,
)
else:
self.model = StableDiffusionInpaintPipeline.from_pretrained(
self.model_id_or_path,
revision="fp16" if use_gpu and fp16 else "main",
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
"""
2022-12-04 06:41:48 +01:00
scheduler_config = self.model.scheduler.config
2023-03-19 15:40:23 +01:00
scheduler = get_scheduler(config.sd_sampler, scheduler_config)
2022-09-22 06:38:32 +02:00
self.model.scheduler = scheduler
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,
guidance_scale=config.sd_guidance_scale,
output_type="np.array",
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),
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
2022-11-25 02:29:20 +01:00
def forward_post_process(self, result, image, mask, config):
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)
return result, image, mask
2022-09-15 16:21:27 +02:00
@staticmethod
def is_downloaded() -> bool:
# model will be downloaded when app start, and can't switch in frontend settings
return True
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"