setup.py

from PIL import Image
import numpy as np
import os
import torch
from datetime import datetime
import time
import collections


from utils import init_weight_dtype, resize_and_crop, resize_and_padding
from model.pipeline import CatVTONPipeline
from model.cloth_masker import AutoMasker, vis_mask
from diffusers.image_processor import VaeImageProcessor
from huggingface_hub import snapshot_download
# torch.backends.cuda.enable_mem_efficient_sdp(False)
# torch.backends.cuda.enable_flash_sdp(False)

def get_files(folder_path, extensions=['py', 'png', 'JPEG']):
    if isinstance(extensions, str):
        extensions = [extensions]
    else:
        extensions = [ex.lower() for ex in extensions]
    result = [x for x in os.listdir(folder_path) if x.split('.')[-1].lower() in extensions]
    return result

base_model_path='booksforcharlie/stable-diffusion-inpainting'
allow_tf32=True
mixed_precision='bf16'
resume_path='zhengchong/CatVTON'
tmp_folder = "/workspace/rs"

automasker = AutoMasker(
    densepose_ckpt=os.path.join(repo_path, "DensePose"),
    schp_ckpt=os.path.join(repo_path, "SCHP"),
    device='cuda',
)

pipeline = CatVTONPipeline(base_ckpt=base_model_path,
                           attn_ckpt=repo_path,
                           attn_ckpt_version="mix",
                           weight_dtype=init_weight_dtype(mixed_precision),
                           use_tf32=allow_tf32,
                           device='cuda')

mask_processor = VaeImageProcessor(vae_scale_factor=8, do_normalize=False, do_binarize=True, do_convert_grayscale=True)

def image_grid(imgs, rows, cols):
    assert len(imgs) == rows * cols

    w, h = imgs[0].size
    grid = Image.new("RGB", size=(cols * w, rows * h))

    for i, img in enumerate(imgs):
        grid.paste(img, box=(i % cols * w, i // cols * h))
    return grid

def inference(
    person_image,
    mask_image,
    cloth_image,
    cloth_type,
    image_size=(1024, 768),
    num_inference_steps=50,
    guidance_scale=2.5,
    seed=42,
    show_type="result only"
):
    start_time = time.time()
    height, width = image_size

    if len(np.unique(np.array(mask_image))) == 1:
        mask_image = None
    else:
        mask_image = np.array(mask_image)
        mask_image[mask_image > 0] = 255
        mask_image = Image.fromarray(mask_image)

    date_str = datetime.now().strftime("%Y%m%d%H%M%S")
    result_save_path = os.path.join(tmp_folder, date_str[:8], date_str[8:] + ".png")
    if not os.path.exists(os.path.join(tmp_folder, date_str[:8])):
        os.makedirs(os.path.join(tmp_folder, date_str[:8]))

    generator = None
    if seed != -1:
        generator = torch.Generator(device='cuda').manual_seed(seed)

    person_image = resize_and_crop(person_image, (width, height))
    cloth_image = resize_and_padding(cloth_image, (width, height))

    # Process mask
    if mask_image is not None:
        mask_image = resize_and_crop(mask_image, (width, height))
    else:
        mask_image = automasker(
            person_image,
            cloth_type
        )['mask']

    mask_image = mask_processor.blur(mask_image, blur_factor=9)

    result_image = pipeline(
        image=person_image,
        condition_image=cloth_image,
        mask=mask_image,
        num_inference_steps=num_inference_steps,
        guidance_scale=guidance_scale,
        generator=generator
    )[0]

    print("FPS: ", 1.0 / (time.time() - start_time))

    # Post-process
    masked_person = vis_mask(person_image, mask_image)
    save_result_image = image_grid([person_image, masked_person, cloth_image, result_image], 1, 4)
    save_result_image.save(result_save_path)

    if show_type == "result only":
        return result_image
    else:
        width, height = person_image.size
        if show_type == "input & result":
            condition_width = width // 2
            conditions = image_grid([person_image, cloth_image], 2, 1)
        else:
            condition_width = width // 3
            conditions = image_grid([person_image, masked_person , cloth_image], 3, 1)
        conditions = conditions.resize((condition_width, height), Image.NEAREST)
        new_result_image = Image.new("RGB", (width + condition_width + 5, height))
        new_result_image.paste(conditions, (0, 0))
        new_result_image.paste(result_image, (condition_width + 5, 0))
    return new_result_image

person_path         = '/workspace/data/person'
mask_path           = None
cloth_path          = '/workspace/data/cloth'
result_path         = '/workspace/data/result'


if not os.path.isfile(person_path):
    os.makedirs(person_path, exist_ok=True)
    person_files = get_files(person_path, extensions=['png', 'jpeg', 'jpg', 'webp'])

    if mask_path:
        os.makedirs(mask_path, exist_ok=True)
        mask_files = [os.path.join(mask_path, f'{os.path.splitext(pf)[0]}.png') for pf in person_files]
    else:
        mask_files = [mask_path] * len(person_files)
    person_files = [os.path.join(person_path, pf) for pf in person_files] if person_files else []
else:
    person_files = [person_path]
    mask_files = [mask_path] * len(person_files)


if not os.path.isfile(cloth_path):
    os.makedirs(cloth_path, exist_ok=True)
    cloth_files = get_files(cloth_path, extensions=['png', 'jpeg', 'jpg', 'webp'])
    cloth_files = [os.path.join(cloth_path, cf) for cf in cloth_files] if cloth_files else []
else:
    cloth_files = [cloth_path]


if not os.path.isdir(result_path):
    os.makedirs(result_path, exist_ok=True)

repo_path = snapshot_download(repo_id=resume_path)

cloth_type          = "upper"
image_size          = (1024, 768)
num_inference_steps = 50
guidance_scale      = 2.5
seed                = 42
show_type           = "all"


for person_file, mask_file in zip(person_files, mask_files):
    for cloth_file in cloth_files:
        person_instance = Image.open(person_file).convert("RGB")
        mask_instance = Image.open(mask_file).convert("L") if mask_file else None
        cloth_instance = Image.open(cloth_file).convert("RGB")

        vton_img = inference(person_instance,
                             mask_instance,
                             cloth_instance,
                             cloth_type,
                             image_size,
                             num_inference_steps,
                             guidance_scale,
                             seed,
                             show_type)
        vton_img.save(os.path.join(result_path, f'{datetime.now().strftime("%Y%m%d%M%S")}.jpg'))