app.py

import spaces
import gradio as gr
import torch
from PIL import Image
from diffusers import DiffusionPipeline, StableDiffusionXLImg2ImgPipeline, AutoencoderKL

device = "cuda"

base_model_id = "stabilityai/stable-diffusion-xl-base-1.0"
vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16)
refiner_id = "stabilityai/stable-diffusion-xl-refiner-1.0"

base_pipeline = DiffusionPipeline.from_pretrained(
    base_model_id, 
    torch_dtype = torch.float16, 
    variant = "fp16", 
    use_safetensors = True
).to(device)

refiner = StableDiffusionXLImg2ImgPipeline.from_pretrained(
    refiner_id, 
    text_encoder_2 = base_pipeline.text_encoder_2, 
    vae = vae,  
    torch_dtype = torch.float16, 
    variant = "fp16", 
    use_safetensors = True
).to(device)



SAMPLER_MAP = {
    "DPM++ Karras SDE": lambda config: DPMSolverMultistepScheduler.from_config(config, use_karras=True, algorithm_type="sde-dpmsolver++"),
    "Euler": lambda config: EulerDiscreteScheduler.from_config(config),
}



@spaces.GPU(duration=59)
def generate(
    prompt, 
    negative_prompt, 
    num_inference_steps, 
    denoising_switch, 
    width, height, 
    guidance_scale
):
    
    base_processed_image = base_pipeline(
        prompt = prompt, 
        negative_prompt = negative_prompt, 
        num_inference_steps = num_inference_steps, 
        denoising_end = denoising_switch, 
        width = width, 
        height = height, 
        guidance_scale = guidance_scale, 
        output_type = "latent"
    ).images

    generated_image = refiner(
        prompt = prompt, 
        negative_prompt = negative_prompt, 
        num_inference_steps = num_inference_steps, 
        denoising_start = denoising_switch, 
        width = width, 
        height = height, 
        guidance_scale = guidance_scale, 
        image = base_processed_image
    ).images[0]

    return generated_image


def create_ui():
    with gr.Blocks() as demo:
        with gr.Row():
            base_model = gr.Radio(label="Base model", choices=[base_model_id], value=base_model_id, interactive=False)
            refiner_model = gr.Radio(label="Refiner model", choices=[refiner_id], value=refiner_id, interactive=False)
        with gr.Row():
            with gr.Column():
                prompt = gr.Textbox(label="Prompt", lines=3)
                negative_prompt = gr.Textbox(label="Negative Prompt", lines=3, value="low quality, bad quality")
                num_inference_steps = gr.Slider(label="Steps", minimum=1, maximum=50, step=1, value=30)
                denoising_switch = gr.Slider(label="Denoising Switch", minimum=0.01, maximum=1, step=0.01, value=0.8)
                width = gr.Slider(label="Width", minimum=64, maximum=2048, step=16, value=1024)
                height = gr.Slider(label="Height", minimum=64, maximum=2048, step=16, value=1024)
                guidance_scale = gr.Slider(label="Guidance Scale", minimum=1.1, maximum=30, step=0.1, value=7.5)
            with gr.Column():
                output_image = gr.Image(interactive=False)
                generate_button = gr.Button("Run", variant="primary")
    
        generate_button.click(
            generate, 
            inputs=[
                prompt, 
                negative_prompt, 
                num_inference_steps, 
                denoising_switch, 
                width, height, 
                guidance_scale
            ], 
            outputs=[output_image]
        )

    return demo


if __name__ == "__main__":
    gradio_app = create_ui()
    gradio_app.launch(
        share = True
    )