app.py

import gradio as gr
from huggingface_hub import login
import os

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()

def infer(use_custom_model, model_name, custom_lora_weight, image_in, prompt, negative_prompt, preprocessor, controlnet_conditioning_scale, guidance_scale, steps, seed, progress=gr.Progress(track_tqdm=True)):
    
    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:
        custom_model = model_name

        # This is where you load your trained weights
        pipe.load_lora_weights(custom_model, use_auth_token=True)
    
    prompt = prompt
    negative_prompt = negative_prompt
    generator = torch.Generator(device="cuda").manual_seed(seed)

    if use_custom_model:
        lora_scale=custom_lora_weight

        images = pipe(
            prompt, 
            negative_prompt=negative_prompt, 
            image=image, 
            controlnet_conditioning_scale=controlnet_conditioning_scale,
            guidance_scale = guidance_scale,
            num_inference_steps=steps,
            generator=generator,
            cross_attention_kwargs={"scale": lora_scale}
        ).images
    else:
        images = pipe(
            prompt, 
            negative_prompt=negative_prompt, 
            image=image, 
            controlnet_conditioning_scale=controlnet_conditioning_scale,
            guidance_scale = guidance_scale,
            num_inference_steps=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;
}
"""
with gr.Blocks(css=css) as demo:
    with gr.Column(elem_id="col-container"):
        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)
                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, type="float")
                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, steps, seed],
        outputs = [result]
    )

demo.queue().launch()