app.py

import torch
import gradio as gr
from diffusers import StableDiffusionXLImg2ImgPipeline
from PIL import Image

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

pipe = StableDiffusionXLImg2ImgPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch.float16 if DEVICE == "cuda" else torch.float32,
    use_safetensors=True,
)

pipe = pipe.to(DEVICE)

# ❌ DO NOT enable cpu offload on HF CPU
# pipe.enable_model_cpu_offload()  <-- REMOVE THIS

def generate(image, prompt):
    image = image.resize((1024, 1024))

    result = pipe(
        prompt=prompt,
        image=image,
        strength=0.2,
        guidance_scale=6,
        num_inference_steps=20,  # keep low for CPU
    )

    return result.images[0]

demo = gr.Interface(
    fn=generate,
    inputs=[
        gr.Image(type="pil", label="Input Image"),
        gr.Textbox(label="Prompt"),
    ],
    outputs=gr.Image(type="pil"),
    title="SDXL Image-to-Image (CPU Safe)",
)

demo.launch()