app.py

import spaces
import torch
from diffusers import DiffusionPipeline
from PIL import Image
import gradio as gr

MODEL_ID = "briaai/BRIA-RMB-3.2"
pipe = DiffusionPipeline.from_pretrained(MODEL_ID, torch_dtype=torch.bfloat16)
pipe.to("cuda")

@spaces.GPU(duration=1500)
def compile_transformer():
    with spaces.aoti_capture(pipe.transformer) as call:
        pipe("arbitrary example prompt")
    
    exported = torch.export.export(
        pipe.transformer,
        args=call.args,
        kwargs=call.kwargs,
    )
    return spaces.aoti_compile(exported)

compiled_transformer = compile_transformer()
spaces.aoti_apply(compiled_transformer, pipe.transformer)

@spaces.GPU(duration=60)
def generate_image(prompt, seed=0):
    torch.manual_seed(seed)
    image = pipe(prompt).images[0]
    return image

with gr.Blocks() as demo:
    gr.Markdown("# BRIA-3.2 Text-to-Image Generator")
    gr.Markdown("Generate images from text prompts using the BRIA-3.2 model.")
    
    with gr.Row():
        prompt = gr.Textbox(
            label="Prompt",
            value="a cat sitting on a chair",
            interactive=True
        )
        seed = gr.Number(
            label="Seed (0 for random)",
            value=0,
            precision=0
        )
    
    generate_btn = gr.Button("Generate Image")
    output = gr.Image(label="Generated Image", type="pil")
    
    generate_btn.click(
        fn=generate_image,
        inputs=[prompt, seed],
        outputs=output
    )
    
    gr.Examples(
        examples=[
            ["a futuristic cityscape at sunset"],
            ["a forest with glowing mushrooms"],
            ["a steampunk robot drinking tea"],
            ["an astronaut riding a horse on mars"]
        ],
        inputs=[prompt],
        outputs=output,
        fn=generate_image,
        cache_examples="lazy"
    )

demo.launch()