app.py

import gradio as gr
import numpy as np
import random

# import spaces #[uncomment to use ZeroGPU]
from diffusers import DiffusionPipeline, AutoPipelineForImage2Image
import torch

device = "cuda" if torch.cuda.is_available() else "cpu"
model_repo_id = "stabilityai/sdxl-turbo"  # Replace to the model you would like to use

if torch.cuda.is_available():
    torch_dtype = torch.float16
else:
    torch_dtype = torch.float32

# Text-to-image pipeline (same as original)
pipe = DiffusionPipeline.from_pretrained(model_repo_id, torch_dtype=torch_dtype)
pipe = pipe.to(device)

# Image-to-image pipeline — built from `pipe` via from_pipe(), so it reuses the
# same unet/vae/text-encoders already in memory. No second checkpoint load,
# no extra VRAM usage.
pipe_img2img = AutoPipelineForImage2Image.from_pipe(pipe)

MAX_SEED = np.iinfo(np.int32).max
MAX_IMAGE_SIZE = 1024


# @spaces.GPU #[uncomment to use ZeroGPU]
def infer(
    prompt,
    init_image,
    negative_prompt,
    seed,
    randomize_seed,
    width,
    height,
    guidance_scale,
    num_inference_steps,
    strength,
    progress=gr.Progress(track_tqdm=True),
):
    if randomize_seed:
        seed = random.randint(0, MAX_SEED)

    generator = torch.Generator().manual_seed(seed)

    if init_image is not None:
        # HYBRID MODE: prompt + reference image -> image-to-image
        init_image = init_image.convert("RGB").resize((width, height))

        # SDXL Turbo requirement: num_inference_steps * strength must be >= 1,
        # otherwise the pipeline errors out. Auto-bump steps if needed instead
        # of crashing.
        steps = num_inference_steps
        if steps * strength < 1:
            steps = max(1, int(np.ceil(1 / max(strength, 1e-3))))
            gr.Warning(
                f"Steps bumped to {steps} so that steps × strength ≥ 1 (SDXL Turbo requirement)."
            )

        image = pipe_img2img(
            prompt=prompt,
            negative_prompt=negative_prompt,
            image=init_image,
            strength=strength,
            guidance_scale=guidance_scale,
            num_inference_steps=steps,
            generator=generator,
        ).images[0]
    else:
        # ORIGINAL MODE: prompt only -> text-to-image
        image = pipe(
            prompt=prompt,
            negative_prompt=negative_prompt,
            guidance_scale=guidance_scale,
            num_inference_steps=num_inference_steps,
            width=width,
            height=height,
            generator=generator,
        ).images[0]

    return image, seed


examples = [
    "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
    "An astronaut riding a green horse",
    "A delicious ceviche cheesecake slice",
]

css = """
#col-container {
    margin: 0 auto;
    max-width: 640px;
}
"""

with gr.Blocks(css=css) as demo:
    with gr.Column(elem_id="col-container"):
        gr.Markdown(" # Text-to-Image + Image-to-Image (Hybrid)")
        gr.Markdown(
            "Upload a reference image below to guide generation with it (image-to-image). "
            "Leave it empty to generate from the prompt alone (text-to-image)."
        )

        with gr.Row():
            prompt = gr.Text(
                label="Prompt",
                show_label=False,
                max_lines=1,
                placeholder="Enter your prompt",
                container=False,
            )

            run_button = gr.Button("Run", scale=0, variant="primary")

        with gr.Row():
            init_image = gr.Image(label="Reference image (optional)", type="pil")
            result = gr.Image(label="Result", show_label=False)

        with gr.Accordion("Advanced Settings", open=False):
            negative_prompt = gr.Text(
                label="Negative prompt",
                max_lines=1,
                placeholder="Enter a negative prompt",
                visible=False,
            )

            seed = gr.Slider(
                label="Seed",
                minimum=0,
                maximum=MAX_SEED,
                step=1,
                value=0,
            )

            randomize_seed = gr.Checkbox(label="Randomize seed", value=True)

            with gr.Row():
                width = gr.Slider(
                    label="Width",
                    minimum=256,
                    maximum=MAX_IMAGE_SIZE,
                    step=32,
                    value=1024,  # Replace with defaults that work for your model
                )

                height = gr.Slider(
                    label="Height",
                    minimum=256,
                    maximum=MAX_IMAGE_SIZE,
                    step=32,
                    value=1024,  # Replace with defaults that work for your model
                )

            with gr.Row():
                guidance_scale = gr.Slider(
                    label="Guidance scale",
                    minimum=0.0,
                    maximum=10.0,
                    step=0.1,
                    value=0.0,  # Replace with defaults that work for your model
                )

                num_inference_steps = gr.Slider(
                    label="Number of inference steps",
                    minimum=1,
                    maximum=50,
                    step=1,
                    value=2,  # Replace with defaults that work for your model
                )

            strength = gr.Slider(
                label="Strength (image-to-image only — higher = further from reference image)",
                minimum=0.0,
                maximum=1.0,
                step=0.05,
                value=0.5,
            )

        gr.Examples(examples=examples, inputs=[prompt])
    gr.on(
        triggers=[run_button.click, prompt.submit],
        fn=infer,
        inputs=[
            prompt,
            init_image,
            negative_prompt,
            seed,
            randomize_seed,
            width,
            height,
            guidance_scale,
            num_inference_steps,
            strength,
        ],
        outputs=[result, seed],
    )

if __name__ == "__main__":
    demo.launch()