app.py

import sys
sys.path.append('./')

import torch
import random
import spaces
import gradio as gr

from diffusers import AutoPipelineForText2Image
from diffusers.utils import load_image

# global variable
device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.float16 if str(device).__contains__("cuda") else torch.float32

def randomize_seed_fn(seed: int, randomize_seed: bool) -> int:
    if randomize_seed:
        seed = random.randint(0, 2000)
    return seed

pipeline = AutoPipelineForText2Image.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=dtype).to(device)
pipeline.load_ip_adapter("h94/IP-Adapter", subfolder="sdxl_models", weight_name="ip-adapter_sdxl.bin")

@spaces.GPU(enable_queue=True)
def create_image(image_pil,
                 prompt,
                 n_prompt,
                 scale, 
                 control_scale, 
                 guidance_scale,
                 num_inference_steps,
                 seed,
                 target="Load only style blocks",
                 ):

    
    if target !="Load original IP-Adapter":
        if target=="Load only style blocks":
            scale = {
                "up": {"block_0": [0.0, control_scale, 0.0]},
            }
        elif target=="Load only layout blocks":
            scale = {
                "down": {"block_2": [0.0, control_scale]},
            }
        elif target == "Load style+layout block":
            scale = {
                "down": {"block_2": [0.0, control_scale]},
                "up": {"block_0": [0.0, control_scale, 0.0]},
            }
        pipeline.set_ip_adapter_scale(scale)
    
    print(image_pil)
    
    style_image = load_image(image_pil)

    generator = torch.Generator(device=device).manual_seed(randomize_seed_fn(seed, False))
    image = pipeline(
        prompt=prompt,
        ip_adapter_image=style_image,
        negative_prompt=n_prompt,
        guidance_scale=guidance_scale,
        num_inference_steps=num_inference_steps,
        generator=generator,
    )
    return image
    

# Description
title = r"""
<h1 align="center">InstantStyle</h1>
"""

description = r"""
How to use:<br>
1. Upload a style image.
2. Set stylization mode, only use style block by default.
2. Enter a text prompt, as done in normal text-to-image models.
3. Click the <b>Submit</b> button to begin customization.
4. Share your stylized photo with your friends and enjoy! 😊


Advanced usage:<br>
1. Click advanced options.
2. Upload another source image for image-based stylization using ControlNet.
3. Enter negative content prompt to avoid content leakage.
"""

article = r"""
---
```bibtex
@article{wang2024instantstyle,
  title={InstantStyle: Free Lunch towards Style-Preserving in Text-to-Image Generation},
  author={Wang, Haofan and Wang, Qixun and Bai, Xu and Qin, Zekui and Chen, Anthony},
  journal={arXiv preprint arXiv:2404.02733},
  year={2024}
}
```
"""

block = gr.Blocks().queue(max_size=10, api_open=True)
with block:
    
    # description
    gr.Markdown(title)
    gr.Markdown(description)
    
    with gr.Tabs():
        with gr.Row():
            with gr.Column():
                
                with gr.Row():
                    with gr.Column():
                        image_pil = gr.Image(label="Style Image", type="pil")
                
                target = gr.Radio(["Load only style blocks", "Load only layout blocks","Load style+layout block", "Load original IP-Adapter"], 
                                  value="Load only style blocks",
                                  label="Style mode")
                
                prompt = gr.Textbox(label="Prompt",
                                    value="a cat, masterpiece, best quality, high quality")
                
                scale = gr.Slider(minimum=0,maximum=2.0, step=0.01,value=1.0, label="Scale")
                
                with gr.Accordion(open=False, label="Advanced Options"):
                    
                    control_scale = gr.Slider(minimum=0,maximum=1.0, step=0.01,value=0.5, label="Controlnet conditioning scale")
                    
                    n_prompt = gr.Textbox(label="Neg Prompt", value="text, watermark, lowres, low quality, worst quality, deformed, glitch, low contrast, noisy, saturation, blurry")
                    guidance_scale = gr.Slider(minimum=1,maximum=15.0, step=0.01,value=5.0, label="guidance scale")
                    num_inference_steps = gr.Slider(minimum=5,maximum=50.0, step=1.0,value=20, label="num inference steps")
                    seed = gr.Slider(minimum=-1000000,maximum=1000000,value=1, step=1, label="Seed Value")
                    randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
                    
                generate_button = gr.Button("Generate Image")
                
            with gr.Column():
                generated_image = gr.Gallery(label="Generated Image")

        generate_button.click(
            fn=randomize_seed_fn,
            inputs=[seed, randomize_seed],
            outputs=seed,
            queue=False,
            api_name=False,
        ).then(
            fn=create_image,
            inputs=[image_pil,
                    prompt,
                    n_prompt,
                    scale, 
                    control_scale, 
                    guidance_scale,
                    num_inference_steps,
                    seed,
                    target], 
            outputs=[generated_image])
    
    gr.Markdown(article)

block.launch(show_error=True)