app_img2img.py

import spaces
import gradio as gr
import time
import torch

from PIL import Image
from segment_utils import(
    segment_image,
    restore_result,
)
from diffusers import (
    StableDiffusionXLImg2ImgPipeline
)

BASE_MODEL = "stabilityai/stable-diffusion-xl-base-1.0"
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

DEFAULT_EDIT_PROMPT = "a beautiful hollywood woman,photo,detailed,8k,high quality,highly detailed,high resolution"
DEFAULT_NEGATIVE_PROMPT = "nude, nudity, nsfw, nipple, Bare-chested, palm hand, hands, fingers, deformed iris, deformed pupils, semi-realistic, cgi, 3d, render, sketch, cartoon, drawing, text, close up, cropped, out of frame, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, cloned face, disfigured"

DEFAULT_CATEGORY = "face"

basepipeline = StableDiffusionXLImg2ImgPipeline.from_pretrained(
    BASE_MODEL,
    torch_dtype=torch.float16,
    variant="fp16",
    use_safetensors=True,
)

basepipeline = basepipeline.to(DEVICE)


@spaces.GPU(duration=15)
def image_to_image(
    input_image: Image,
    edit_prompt: str,
    seed: int,
    num_steps: int,
    guidance_scale: float,
):
    run_task_time = 0
    time_cost_str = ''
    run_task_time, time_cost_str = get_time_cost(run_task_time, time_cost_str)
    generator = torch.Generator(device=DEVICE).manual_seed(seed)
    generated_image = basepipeline(
        generator=generator,
        prompt=edit_prompt,
        negative_prompt=DEFAULT_NEGATIVE_PROMPT,
        image=input_image,
        guidance_scale=guidance_scale,
        num_inference_steps = num_steps,
    ).images[0]
    
    run_task_time, time_cost_str = get_time_cost(run_task_time, time_cost_str)

    return generated_image, time_cost_str

def get_time_cost(run_task_time, time_cost_str):
    now_time = int(time.time()*1000)
    if run_task_time == 0:
        time_cost_str = 'start'
    else:
        if time_cost_str != '': 
            time_cost_str += f'-->'
        time_cost_str += f'{now_time - run_task_time}'
    run_task_time = now_time
    return run_task_time, time_cost_str

def create_demo() -> gr.Blocks:
    with gr.Blocks() as demo:
        croper = gr.State()
        with gr.Row():
            with gr.Column():
                edit_prompt = gr.Textbox(lines=1, label="Edit Prompt", value=DEFAULT_EDIT_PROMPT)
                generate_size = gr.Number(label="Generate Size", value=1024)
                category = gr.Textbox(label="Category", value=DEFAULT_CATEGORY, visible=False)
            with gr.Column():
                num_steps = gr.Slider(minimum=1, maximum=100, value=30, step=1, label="Num Steps")
                guidance_scale = gr.Slider(minimum=0, maximum=30, value=15, step=0.5, label="Guidance Scale")
                mask_expansion = gr.Number(label="Mask Expansion", value=300, visible=False)
            with gr.Column():
                mask_dilation = gr.Slider(minimum=0, maximum=10, value=2, step=1, label="Mask Dilation")
                seed = gr.Number(label="Seed", value=8)
                g_btn = gr.Button("Edit Image")
                
        with gr.Row():
            with gr.Column():
                input_image = gr.Image(label="Input Image", type="pil")
            with gr.Column():
                restored_image = gr.Image(label="Restored Image", type="pil", interactive=False)
            with gr.Column():
                origin_area_image = gr.Image(label="Origin Area Image", type="pil", interactive=False)
                generated_image = gr.Image(label="Generated Image", type="pil", interactive=False)
                generated_cost = gr.Textbox(label="Time cost by step (ms):", visible=True, interactive=False)
        
        g_btn.click(
            fn=segment_image,
            inputs=[input_image, category, generate_size, mask_expansion, mask_dilation],
            outputs=[origin_area_image, croper],
        ).success(
            fn=image_to_image,
            inputs=[origin_area_image, edit_prompt,seed, num_steps, guidance_scale],
            outputs=[generated_image, generated_cost],
        ).success(
            fn=restore_result,
            inputs=[croper, category, generated_image],
            outputs=[restored_image],
        )

    return demo