app.py

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

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


@logger.catch(reraise=True)
@spaces.GPU(duration=180)
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))

    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

    if model == "timbrooks/instruct-pix2pix":
        pipe = StableDiffusionInstructPix2PixPipeline.from_pretrained(model).to("cuda")

        images = pipe(
            prompt=prompt,
            image=init_image,
            callback_on_step_end=progress_callback,
        ).images
    else:
        pipe = AutoPipelineForImage2Image.from_pretrained(model).to("cuda")

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


demo = gr.Interface(
    fn=generate,
    inputs=[
        gr.Dropdown(
            label="Model", choices=models, value=models[0], 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()