app.py

import spaces  # MUST be imported before any CUDA-touching package (torch/diffusers)

import gradio as gr
import numpy as np
import random
import torch
from diffusers import Flux2KleinPipeline

# ---------------------------------------------------------------------------
# Model: FLUX.2 [klein] 4B
#   - Apache-2.0, 4B params, BFL's fastest small model (sub-second, ~13GB VRAM)
#   - Unified text-to-image + multi-reference editing
#   - Released Jan 2026 (current BFL small-model generation)
# ---------------------------------------------------------------------------
MODEL_REPO_ID = "black-forest-labs/FLUX.2-klein-4B"

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

# Load on cuda at module level. (No enable_model_cpu_offload() on ZeroGPU —
# the GPU is only attached inside @spaces.GPU; module-level cuda uses the
# ZeroGPU CUDA-emulation, and offload would conflict.)
pipe = Flux2KleinPipeline.from_pretrained(MODEL_REPO_ID, torch_dtype=dtype).to(device)

MAX_SEED = np.iinfo(np.int32).max
MAX_IMAGE_SIZE = 2048


@spaces.GPU(duration=60)
def infer(
    prompt,
    seed,
    randomize_seed,
    width,
    height,
    num_inference_steps,
    guidance_scale,
    progress=gr.Progress(track_tqdm=True),
):
    if randomize_seed:
        seed = random.randint(0, MAX_SEED)
    generator = torch.Generator(device=device).manual_seed(seed)
    image = pipe(
        prompt=prompt,
        width=width,
        height=height,
        num_inference_steps=num_inference_steps,
        guidance_scale=guidance_scale,
        generator=generator,
    ).images[0]
    return image, seed


examples = [
    "A magical city at twilight, glowing windows, storybook illustration, warm light",
    "A cat holding a sign that says hello world",
    "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
]

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("# 🖼️ NEXUS Visual Weaver — FLUX.2 [klein] 4B")
        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")

        result = gr.Image(label="Result", show_label=False)

        with gr.Accordion("Advanced Settings", open=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)
                height = gr.Slider(label="Height", minimum=256, maximum=MAX_IMAGE_SIZE, step=32, value=1024)
            with gr.Row():
                num_inference_steps = gr.Slider(
                    label="Inference steps", minimum=1, maximum=8, step=1, value=4
                )
                guidance_scale = gr.Slider(
                    label="Guidance scale", minimum=0.0, maximum=5.0, step=0.1, value=1.0
                )

        gr.Examples(examples=examples, inputs=[prompt])

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

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