app.py

import torch
import gradio as gr
from diffusers import FluxPipeline
import spaces

from nunchaku import NunchakuFluxTransformer2dModel
from nunchaku.utils import get_precision

dtype=torch.bfloat16
device = "cuda" if torch.cuda.is_available() else "cpu"

@spaces.GPU(duration=10)
def gpu_precision():
    precision = get_precision()
    return precision

transformer = NunchakuFluxTransformer2dModel.from_pretrained(
    f"nunchaku-tech/nunchaku-flux.1-dev/svdq-{gpu_precision()}_r32-flux.1-dev.safetensors"
)
pipeline = FluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-dev", transformer=transformer, torch_dtype=dtype
).to(device)

@spaces.GPU(duration=60)
def generate_image(prompt: str, steps: int, guidance_scale: float):
    if not prompt.strip():
        raise gr.Error("Prompt cannot be empty.")
    
    with torch.inference_mode(), torch.autocast(device, dtype=dtype):
        result = pipeline(
            prompt=prompt,
            width=576,
            height=1024,
            num_inference_steps=steps,
            guidance_scale=guidance_scale
        )
    return result.images[0]

# Minimal Gradio UI
demo = gr.Interface(
    fn=generate_image,
    inputs=[
        gr.Textbox(label="Prompt", placeholder="Describe the scene..."),
        gr.Slider(label="Inference Steps", minimum=5, maximum=50, step=1, value=20),
        gr.Slider(label="Guidance Scale", minimum=0.1, maximum=10.0, step=0.1, value=3.5)
    ],
    outputs=gr.Image(label="Generated Image"),
    title="FLUX Image Generator",
    description="Prompt-based image generation using Flux + LoRA"
)

demo.launch()