app.py

import spaces
import torch
from diffusers import (
    StableDiffusionXLPipeline,
    StableDiffusionXLImg2ImgPipeline
)
from PIL import Image
import gradio as gr

MODEL_ID = "stabilityai/stable-diffusion-xl-base-1.0"
IP_ADAPTER_REPO = "h94/IP-Adapter"

# Load pipelines
pipe_txt2img = StableDiffusionXLPipeline.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.float32
).to("cuda" if torch.cuda.is_available() else "cpu")

pipe_img2img = StableDiffusionXLImg2ImgPipeline.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.float32
).to("cuda" if torch.cuda.is_available() else "cpu")

# Load IP-Adapter
for p in [pipe_txt2img, pipe_img2img]:
    p.load_ip_adapter(
        IP_ADAPTER_REPO,
        subfolder="sdxl_models",
        weight_name="ip-adapter_sdxl.bin"
    )
    p.set_ip_adapter_scale(0.7)

@spaces.GPU
def generate(prompt, init_image, ref_images):
    ref_list = []
    if ref_images:
        for item in ref_images:
            img = item[0] if isinstance(item, tuple) else item
            if img is not None:
                if isinstance(img, str):
                    img = Image.open(img)
                ref_list.append(img.convert("RGB"))

    common_kwargs = {
        "prompt": prompt,
        "ip_adapter_image": [ref_list] if len(ref_list) > 0 else None,
        "num_inference_steps": 30,
        "guidance_scale": 8.0
    }

    # CASE 1: Img2Img
    if init_image is not None:
        init_image = init_image.convert("RGB")

        result = pipe_img2img(
            image=init_image,
            strength=0.5,
            **common_kwargs
        ).images[0]

    # CASE 2: Text2Img
    else:
        result = pipe_txt2img(
            **common_kwargs
        ).images[0]

    return result


with gr.Blocks() as demo:
    gr.Markdown("# SDXL Character Generator (IP-Adapter + Multi-Image + Img2Img)")

    with gr.Row():
        prompt = gr.Textbox(label="Prompt", placeholder="cinematic portrait, ultra detailed, same person")

    with gr.Row():
        init_image = gr.Image(label="Init Image (Img2Img)", type="pil")
        ref_images = gr.Gallery(label="Reference Images (Multi-image)", columns=3, object_fit="contain")

    generate_btn = gr.Button("Generate")

    output = gr.Image(label="Output")

    generate_btn.click(
        fn=generate,
        inputs=[prompt, init_image, ref_images],
        outputs=output
    )


demo.launch()