IOPaint/iopaint/model/kandinsky.py

66 lines
1.9 KiB
Python
Raw Normal View History

2023-08-30 15:30:11 +02:00
import PIL.Image
import cv2
import numpy as np
import torch
2024-01-05 09:38:55 +01:00
from iopaint.const import KANDINSKY22_NAME
2024-01-05 09:40:06 +01:00
from .base import DiffusionInpaintModel
2024-01-05 08:19:23 +01:00
from iopaint.schema import InpaintRequest
2024-01-09 15:42:48 +01:00
from .utils import get_torch_dtype, enable_low_mem
2023-08-30 15:30:11 +02:00
class Kandinsky(DiffusionInpaintModel):
pad_mod = 64
min_size = 512
def init_model(self, device: torch.device, **kwargs):
from diffusers import AutoPipelineForInpainting
use_gpu, torch_dtype = get_torch_dtype(device, kwargs.get("no_half", False))
2023-08-30 15:30:11 +02:00
model_kwargs = {
"torch_dtype": torch_dtype,
}
self.model = AutoPipelineForInpainting.from_pretrained(
2023-12-27 15:00:07 +01:00
self.name, **model_kwargs
2023-08-30 15:30:11 +02:00
).to(device)
2024-01-09 15:42:48 +01:00
enable_low_mem(self.model, kwargs.get("low_mem", False))
2023-08-30 15:30:11 +02:00
self.callback = kwargs.pop("callback", None)
2023-12-30 16:36:44 +01:00
def forward(self, image, mask, config: InpaintRequest):
2023-08-30 15:30:11 +02:00
"""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
"""
2024-01-02 07:34:36 +01:00
self.set_scheduler(config)
2023-08-30 15:30:11 +02:00
generator = torch.manual_seed(config.sd_seed)
mask = mask.astype(np.float32) / 255
img_h, img_w = image.shape[:2]
2023-11-14 07:02:10 +01:00
# kandinsky 没有 strength
2023-08-30 15:30:11 +02:00
output = self.model(
prompt=config.prompt,
negative_prompt=config.negative_prompt,
image=PIL.Image.fromarray(image),
mask_image=mask[:, :, 0],
height=img_h,
width=img_w,
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,
2023-11-14 07:02:10 +01:00
generator=generator,
2023-08-30 15:30:11 +02:00
).images[0]
output = (output * 255).round().astype("uint8")
output = cv2.cvtColor(output, cv2.COLOR_RGB2BGR)
return output
class Kandinsky22(Kandinsky):
2024-01-05 09:38:55 +01:00
name = KANDINSKY22_NAME