app.py

import gradio as gr
import PIL
from PIL import Image
import numpy as np
import os
import uuid
import torch
from torch import autocast
import cv2
from io import BytesIO

from matplotlib import pyplot as plt
from torchvision import transforms

import io
import logging
import multiprocessing
import random
import time
import imghdr
from pathlib import Path
from typing import Union
from loguru import logger

from lama_cleaner.model_manager import ModelManager
from lama_cleaner.schema import Config

try:
    torch._C._jit_override_can_fuse_on_cpu(False)
    torch._C._jit_override_can_fuse_on_gpu(False)
    torch._C._jit_set_texpr_fuser_enabled(False)
    torch._C._jit_set_nvfuser_enabled(False)
except:
    pass


from lama_cleaner.helper import (
    load_img,
    numpy_to_bytes,
    resize_max_size,
)

NUM_THREADS = str(multiprocessing.cpu_count())

# fix libomp problem on windows https://github.com/Sanster/lama-cleaner/issues/56
os.environ["KMP_DUPLICATE_LIB_OK"] = "True"

os.environ["OMP_NUM_THREADS"] = NUM_THREADS
os.environ["OPENBLAS_NUM_THREADS"] = NUM_THREADS
os.environ["MKL_NUM_THREADS"] = NUM_THREADS
os.environ["VECLIB_MAXIMUM_THREADS"] = NUM_THREADS
os.environ["NUMEXPR_NUM_THREADS"] = NUM_THREADS
if os.environ.get("CACHE_DIR"):
    os.environ["TORCH_HOME"] = os.environ["CACHE_DIR"]

HF_TOKEN_SD = os.environ.get('HF_TOKEN_SD')

device = "cuda" if torch.cuda.is_available() else "cpu"
print(f'device = {device}')

def get_image_ext(img_bytes):
    w = imghdr.what("", img_bytes)
    if w is None:
        w = "jpeg"
    return w
    
def read_content(file_path):
    """read the content of target file
    """
    with open(file_path, 'rb') as f:
        content = f.read()
    return content
    
def get_image_enhancer(scale = 2, device='cuda:0'):
    from basicsr.archs.rrdbnet_arch import RRDBNet
    from realesrgan import RealESRGANer
    from realesrgan.archs.srvgg_arch import SRVGGNetCompact
    from gfpgan import GFPGANer

    realesrgan_model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64,
                    num_block=23, num_grow_ch=32, scale=4
                  )
    netscale = scale

    model_realesrgan = 'https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth'
    upsampler = RealESRGANer(
          scale=netscale,
          model_path=model_realesrgan,
          model=realesrgan_model,
          tile=0,
          tile_pad=10,
          pre_pad=0,
          half=False if device=='cpu' else True,
          device=device
    )

    model_GFPGAN = 'https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.3.pth'
    img_enhancer = GFPGANer(
          model_path=model_GFPGAN,
          upscale=scale,
          arch='clean',
          channel_multiplier=2,
          bg_upsampler=upsampler,
          device=device
      )
    return img_enhancer

image_enhancer = get_image_enhancer(scale = 1, device=device)
        
model = None

def model_process(image, mask, img_enhancer):
    global model,image_enhancer
    
    if mask.shape[0] == image.shape[1] and mask.shape[1] == image.shape[0] and mask.shape[0] != mask.shape[1]:
        # rotate image
        image = np.transpose(image[::-1, ...][:, ::-1], axes=(1, 0, 2))[::-1, ...]
        
    original_shape = image.shape
    interpolation = cv2.INTER_CUBIC
    
    size_limit = 1080 #1080 # "Original"
    if size_limit == "Original":
        size_limit = max(image.shape)
    else:
        size_limit = int(size_limit)

    config = Config(
        ldm_steps=25,
        ldm_sampler='plms',
        zits_wireframe=True,
        hd_strategy='Original',
        hd_strategy_crop_margin=196,
        hd_strategy_crop_trigger_size=1280,
        hd_strategy_resize_limit=2048,
        prompt='',
        use_croper=False,
        croper_x=0,
        croper_y=0,
        croper_height=512,
        croper_width=512,
        sd_mask_blur=5,
        sd_strength=0.75,
        sd_steps=50,
        sd_guidance_scale=7.5,
        sd_sampler='ddim',
        sd_seed=42,
        cv2_flag='INPAINT_NS',
        cv2_radius=5,
    )
    
    if config.sd_seed == -1:
        config.sd_seed = random.randint(1, 999999999)

    logger.info(f"Origin image shape_0_: {original_shape} / {size_limit}")
    image = resize_max_size(image, size_limit=size_limit, interpolation=interpolation)
    logger.info(f"Resized image shape_1_: {image.shape}")
    
    logger.info(f"mask image shape_0_: {mask.shape} / {type(mask)}")
    mask = resize_max_size(mask, size_limit=size_limit, interpolation=interpolation)
    logger.info(f"mask image shape_1_: {mask.shape} / {type(mask)}")

    if model is None:
        return None
        
    res_np_img = model(image, mask, config)
    torch.cuda.empty_cache()
  
    image = Image.open(io.BytesIO(numpy_to_bytes(res_np_img, 'png')))
    
    if image_enhancer is not None and img_enhancer:
        start = time.time()
        input_img_rgb = np.array(image)
        input_img_bgr = input_img_rgb[...,[2,1,0]]                        
        _, _, enhance_img = image_enhancer.enhance(input_img_bgr, has_aligned=False,
                                                only_center_face=False, paste_back=True)  
        input_img_rgb = enhance_img[...,[2,1,0]]
        img_enhance = Image.fromarray(np.uint8(input_img_rgb))                                                         
        image = img_enhance  
        log_info = f"image_enhancer_: {(time.time() - start) * 1000}ms, {res_np_img.shape} "
        logger.info(log_info)
         
    return  image # image
    
model = ModelManager(
        name='lama',
        device=device,
    )

image_type = 'pil' # filepath' 
def predict(input, img_enhancer):
    if input is None:
        return None
    if image_type == 'filepath':
        # input: {'image': '/tmp/tmp8mn9xw93.png', 'mask': '/tmp/tmpn5ars4te.png'}
        origin_image_bytes = read_content(input["image"])
        print(f'origin_image_bytes = ', type(origin_image_bytes), len(origin_image_bytes))    
        image, _ = load_img(origin_image_bytes) 
        mask, _ = load_img(read_content(input["mask"]), gray=True)       
    elif image_type == 'pil':
        # input: {'image': pil, 'mask': pil}
        image_pil = input['image']
        mask_pil = input['mask']
        image = np.array(image_pil)
        mask = np.array(mask_pil.convert("L"))
    output = model_process(image, mask, img_enhancer)
    return output

css = '''
.container {max-width: 100%;margin: auto;padding-top: 1.5rem}
.output-image, .input-image, .image-preview {height: 600px !important;object-fit: contain}
#work-container {min-width: min(160px, 100%) !important;flex-grow: 0 !important}
#image_upload{min-height:610px}
#image_upload [data-testid="image"], #image_upload [data-testid="image"] > div{min-height: 620px}
#image_output{margin: 0 auto; text-align: center;width:640px}
#erase-container{margin: 0 auto; text-align: center;width:150px;border-width:5px;border-color:#2c9748}
#enhancer-checkbox{width:520px}
#enhancer-tip{width:450px}
#enhancer-tip-div{text-align: left}
#prompt-container{margin: 0 auto; text-align: center;width:fit-content;min-width: min(150px, 100%);flex-grow: 0; flex-wrap: nowrap;}
.footer {margin-bottom: 45px;margin-top: 35px;text-align: center;border-bottom: 1px solid #e5e5e5}
.footer>p {font-size: .8rem; display: inline-block; padding: 0 10px;transform: translateY(10px);background: white}
.dark .footer {border-color: #303030}
.dark .footer>p {background: #0b0f19}
#image_upload .touch-none{display: flex}
@keyframes spin {
    from {
        transform: rotate(0deg);
    }
    to {
        transform: rotate(360deg);
    }
}
'''

image_blocks = gr.Blocks(css=css)
with image_blocks as demo:
    with gr.Group():
        with gr.Box(elem_id="work-container"):
            with gr.Row(elem_id="input-container"):
                with gr.Column():
                    image = gr.Image(source='upload', elem_id="image_upload",tool='sketch', type=f'{image_type}', 
                                     label="Upload(载入图片)", show_label=True).style(mobile_collapse=False)
            with gr.Row(elem_id="prompt-container").style(mobile_collapse=False, equal_height=True):
                with gr.Column(elem_id="erase-container"):
                    btn_erase = gr.Button(value = "Erase(擦除↓)",elem_id="erase_btn").style(
                        margin=True,
                        rounded=(True, True, True, True),
                        full_width=True,
                    ).style(width=100)   
                with gr.Column(elem_id="enhancer-checkbox"):
                    enhancer_label = 'Enhanced image(processing is very slow, please check only for blurred images)【增强图像(处理很慢，请仅针对模糊图像做勾选)】'
                    img_enhancer = gr.Checkbox(label=enhancer_label).style(width=150) 
            with gr.Row(elem_id="output-container"):             
                with gr.Column():
                    image_out = gr.Image(label="Result", elem_id="image_output", visible=True).style(width=640)
                    
            btn_erase.click(fn=predict, inputs=[image, img_enhancer], outputs=[image_out])
            
image_blocks.launch()