app.py

import gradio as gr
from huggingface_hub import login
import os

is_shared_ui = True if "fffiloni/sdxl-control-loras" in os.environ['SPACE_ID'] else False

hf_token = os.environ.get("HF_TOKEN")
login(token=hf_token)

from diffusers import ControlNetModel, StableDiffusionXLControlNetPipeline, AutoencoderKL
from diffusers.utils import load_image
from PIL import Image
import torch
import numpy as np
import cv2

vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16)

controlnet = ControlNetModel.from_pretrained(
    "diffusers/controlnet-canny-sdxl-1.0",
    torch_dtype=torch.float16
)

pipe = StableDiffusionXLControlNetPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    controlnet=controlnet,
    vae=vae,
    torch_dtype=torch.float16, 
    variant="fp16",
    use_safetensors=True
)

pipe.to("cuda")



#pipe.enable_model_cpu_offload()

from PIL import Image

def resize_image(input_path, output_path, target_height):
    # Open the input image
    img = Image.open(input_path)

    # Calculate the aspect ratio of the original image
    original_width, original_height = img.size
    original_aspect_ratio = original_width / original_height

    # Calculate the new width while maintaining the aspect ratio and the target height
    new_width = int(target_height * original_aspect_ratio)

    # Resize the image while maintaining the aspect ratio and fixing the height
    img = img.resize((new_width, target_height), Image.LANCZOS)

    # Save the resized image
    img.save(output_path)

    return output_path

def infer(use_custom_model, model_name, custom_lora_weight, image_in, prompt, negative_prompt, preprocessor, controlnet_conditioning_scale, guidance_scale, inf_steps, seed, progress=gr.Progress(track_tqdm=True)):
    prompt = prompt
    negative_prompt = negative_prompt
    generator = torch.Generator(device="cuda").manual_seed(seed)

    if image_in == None:
        raise gr.Error("You forgot to upload a source image.")
    
    image_in = resize_image(image_in, "resized_input.jpg", 1024)
    
    if preprocessor == "canny":

        image = load_image(image_in)

        image = np.array(image)
        image = cv2.Canny(image, 100, 200)
        image = image[:, :, None]
        image = np.concatenate([image, image, image], axis=2)
        image = Image.fromarray(image)
    
    if use_custom_model:
        
        if custom_model == "":
            raise gr.Error("you forgot to set a custom model name.")
        
        custom_model = model_name

        # This is where you load your trained weights
        pipe.load_lora_weights(custom_model, use_auth_token=True)
    
        lora_scale=custom_lora_weight

        images = pipe(
            prompt, 
            negative_prompt=negative_prompt, 
            image=image, 
            controlnet_conditioning_scale=float(controlnet_conditioning_scale),
            guidance_scale = float(guidance_scale),
            num_inference_steps=inf_steps,
            generator=generator,
            cross_attention_kwargs={"scale": lora_scale}
        ).images
    else:
        images = pipe(
            prompt, 
            negative_prompt=negative_prompt, 
            image=image, 
            controlnet_conditioning_scale=float(controlnet_conditioning_scale),
            guidance_scale = float(guidance_scale),
            num_inference_steps=inf_steps,
            generator=generator,
        ).images

    images[0].save(f"result.png")

    return f"result.png"

css="""
#col-container{
    margin: 0 auto;
    max-width: 680px;
    text-align: left;
}
div#warning-duplicate {
    background-color: #ebf5ff;
    padding: 0 10px 5px;
    margin: 20px 0;
}
div#warning-duplicate > .gr-prose > h2, div#warning-duplicate > .gr-prose > p {
    color: #0f4592!important;
}
div#warning-duplicate strong {
    color: #0f4592;
}
p.actions {
    display: flex;
    align-items: center;
    margin: 20px 0;
}
div#warning-duplicate .actions a {
    display: inline-block;
    margin-right: 10px;
}
"""

with gr.Blocks(css=css) as demo:
    with gr.Column(elem_id="col-container"):
        if is_shared_ui:
            top_description = gr.HTML(f'''
                <div class="gr-prose">
                    <h2><svg xmlns="http://www.w3.org/2000/svg" width="18px" height="18px" style="margin-right: 0px;display: inline-block;"fill="none"><path fill="#fff" d="M7 13.2a6.3 6.3 0 0 0 4.4-10.7A6.3 6.3 0 0 0 .6 6.9 6.3 6.3 0 0 0 7 13.2Z"/><path fill="#fff" fill-rule="evenodd" d="M7 0a6.9 6.9 0 0 1 4.8 11.8A6.9 6.9 0 0 1 0 7 6.9 6.9 0 0 1 7 0Zm0 0v.7V0ZM0 7h.6H0Zm7 6.8v-.6.6ZM13.7 7h-.6.6ZM9.1 1.7c-.7-.3-1.4-.4-2.2-.4a5.6 5.6 0 0 0-4 1.6 5.6 5.6 0 0 0-1.6 4 5.6 5.6 0 0 0 1.6 4 5.6 5.6 0 0 0 4 1.7 5.6 5.6 0 0 0 4-1.7 5.6 5.6 0 0 0 1.7-4 5.6 5.6 0 0 0-1.7-4c-.5-.5-1.1-.9-1.8-1.2Z" clip-rule="evenodd"/><path fill="#000" fill-rule="evenodd" d="M7 2.9a.8.8 0 1 1 0 1.5A.8.8 0 0 1 7 3ZM5.8 5.7c0-.4.3-.6.6-.6h.7c.3 0 .6.2.6.6v3.7h.5a.6.6 0 0 1 0 1.3H6a.6.6 0 0 1 0-1.3h.4v-3a.6.6 0 0 1-.6-.7Z" clip-rule="evenodd"/></svg>
                    Note: you might want to use a custom LoRa model</h2>
                    <p class="main-message">
                        To do so, <strong>duplicate the Space</strong> and run it on your own profile using <strong>your own access token</strong> and eventually a GPU (T4-small or A10G-small) for faster inference without waiting in the queue.<br />
                    </p>
                    <p class="actions">
                        <a href="https://huggingface.co/spaces/{os.environ['SPACE_ID']}?duplicate=true">
                            <img src="https://huggingface.co/datasets/huggingface/badges/resolve/main/duplicate-this-space-lg-dark.svg" alt="Duplicate this Space" />
                        </a>
                        to start using private models and skip the queue
                    </p>
                </div>
            ''', elem_id="warning-duplicate")
        gr.HTML("""
<h2 style="text-align: center;">SD-XL Control LoRas</h2>
<p style="text-align: center;">Use StableDiffusion XL with <a href="https://huggingface.co/collections/diffusers/sdxl-controlnets-64f9c35846f3f06f5abe351f">Diffusers' SDXL ControlNets</a></p>

        """)
        
        image_in = gr.Image(source="upload", type="filepath")
        with gr.Row():
            with gr.Column():
                prompt = gr.Textbox(label="Prompt")
                negative_prompt = gr.Textbox(label="Negative prompt", value="extra digit, fewer digits, cropped, worst quality, low quality, glitch, deformed, mutated, ugly, disfigured")
                guidance_scale = gr.Slider(label="Guidance Scale", minimum=1.0, maximum=10.0, step=0.1, value=7.5)
                inf_steps = gr.Slider(label="Inference Steps", minimum="25", maximum="50", step=1, value=25)
            with gr.Column():
                preprocessor = gr.Dropdown(label="Preprocessor", choices=["canny"], value="canny", interactive=False, info="For the moment, only canny is available")  
                controlnet_conditioning_scale = gr.Slider(label="Controlnet conditioning Scale", minimum=0.1, maximum=0.9, step=0.01, value=0.5)
                seed = gr.Slider(label="seed", minimum=0, maximum=500000, step=1, value=42)
        use_custom_model = gr.Checkbox(label="Use a public custom model ?(optional)", value=False, info="To use a private model, you'll prefer to duplicate the space with your own access token.")
        with gr.Row():
            model_name = gr.Textbox(label="Custom Model to use", placeholder="username/my_custom_public_model")
            custom_lora_weight = gr.Slider(label="Custom model weights", minimum=0.1, maximum=0.9, step=0.1, value=0.9)
        submit_btn = gr.Button("Submit")
        result = gr.Image(label="Result")

    submit_btn.click(
        fn = infer,
        inputs = [use_custom_model, model_name, custom_lora_weight, image_in, prompt, negative_prompt, preprocessor, controlnet_conditioning_scale, guidance_scale, inf_steps, seed],
        outputs = [result]
    )

demo.queue(max_size=12).launch()