app.py

import gradio as gr
import spaces
from diffusers import AutoPipelineForImage2Image, StableDiffusionInstructPix2PixPipeline
from loguru import logger
from PIL import Image

models = [
    "stabilityai/sdxl-turbo",
    "stabilityai/stable-diffusion-3-medium-diffusers",
    "stabilityai/stable-diffusion-xl-refiner-1.0",
    "timbrooks/instruct-pix2pix",
]
DEFAULT_MODEL = "stabilityai/stable-diffusion-xl-refiner-1.0"


def load_pipeline(model):
    pipeline_type = (
        StableDiffusionInstructPix2PixPipeline
        if model == "timbrooks/instruct-pix2pix"
        else AutoPipelineForImage2Image
    )

    return pipeline_type.from_pretrained(model)


load_pipeline(DEFAULT_MODEL).to("cuda")
loaded_models = {DEFAULT_MODEL}


def generate_image(
    model: str,
    prompt: str,
    init_image: Image.Image,
    strength: float,
    progress,
):
    logger.debug(f"Loading pipeline: {dict(model=model)}")
    pipe = load_pipeline(model).to("cuda")

    logger.debug(f"Generating image: {dict(prompt=prompt)}")
    additional_args = (
        {} if model == "timbrooks/instruct-pix2pix" else dict(strength=strength)
    )

    def progress_callback(pipe, step_index, timestep, callback_kwargs):
        logger.trace(
            f"Callback: {dict(num_timesteps=pipe.num_timesteps, step_index=step_index, timestep=timestep)}"
        )
        progress((step_index + 1, pipe.num_timesteps))
        return callback_kwargs

    images = pipe(
        prompt=prompt,
        image=init_image,
        callback_on_step_end=progress_callback,
        **additional_args,
    ).images
    return images[0]


@spaces.GPU
def gpu(*args, **kwargs):
    return generate_image(*args, **kwargs)


@spaces.GPU(duration=180)
def gpu_3min(*args, **kwargs):
    return generate_image(*args, **kwargs)


@logger.catch(reraise=True)
def generate(
    model: str,
    prompt: str,
    init_image: Image.Image,
    strength: float,
    progress=gr.Progress(),
):
    logger.info(
        f"Starting image generation: {dict(model=model, prompt=prompt, image=init_image, strength=strength)}"
    )

    # Downscale the image
    init_image.thumbnail((1024, 1024))

    # Cache the model files for the pipeline
    if model not in loaded_models:
        logger.debug(f"Caching pipeline: {dict(model=model)}")

        load_pipeline(model)
        loaded_models.add(model)

    gpu_runner = gpu_3min if model == "timbrooks/instruct-pix2pix" else gpu

    return gpu_runner(model, prompt, init_image, strength, progress)


demo = gr.Interface(
    fn=generate,
    inputs=[
        gr.Dropdown(
            label="Model", choices=models, value=DEFAULT_MODEL, allow_custom_value=True
        ),
        gr.Text(label="Prompt"),
        gr.Image(label="Init image", type="pil"),
        gr.Slider(label="Strength", minimum=0, maximum=1, value=0.3),
    ],
    outputs=[gr.Image(label="Output")],
)

demo.launch()