app.py

import gradio as gr
import numpy as np
import spaces
import torch
import random
import os
from PIL import Image

# Import the pipeline from diffusers
from diffusers import FluxKontextPipeline

# --- Constants ---
MAX_SEED = np.iinfo(np.int32).max

# --- Global pipeline variable ---
pipe = None

def load_model():
    """Load the model on CPU first, then move to GPU when needed"""
    global pipe
    if pipe is None:
        # Get token from environment variable
        hf_token = os.getenv("HF_TOKEN")
        if hf_token:
            pipe = FluxKontextPipeline.from_pretrained(
                "black-forest-labs/FLUX.1-Kontext-dev", 
                torch_dtype=torch.bfloat16,
                token=hf_token,
            )
        else:
            raise gr.Error("HF_TOKEN environment variable not found. Please add your Hugging Face token to the Space settings.")
    return pipe

# --- Core Inference Function for ChatInterface ---
@spaces.GPU(duration=120)  # Set duration based on expected inference time
def chat_fn(message, chat_history, seed, randomize_seed, guidance_scale, steps, progress=gr.Progress(track_tqdm=True)):
    """
    Performs image generation or editing based on user input from the chat interface.
    """
    # Load and move model to GPU within the decorated function
    pipe = load_model()
    pipe = pipe.to("cuda")
    
    prompt = message["text"]
    files = message["files"]

    if not prompt and not files:
        raise gr.Error("Please provide a prompt and/or upload an image.")

    if randomize_seed:
        seed = random.randint(0, MAX_SEED)

    generator = torch.Generator(device="cuda").manual_seed(int(seed))

    input_image = None
    if files:
        print(f"Received image: {files[0]}")
        input_image = Image.open(files[0]).convert("RGB")
        image = pipe(
            image=input_image,
            prompt=prompt,
            guidance_scale=guidance_scale,
            num_inference_steps=steps,
            generator=generator,
        ).images[0]
    else:
        print(f"Received prompt for text-to-image: {prompt}")
        image = pipe(
            prompt=prompt,
            guidance_scale=guidance_scale,
            num_inference_steps=steps,
            generator=generator,
        ).images[0]
    
    # Move model back to CPU to free GPU memory
    pipe = pipe.to("cpu")
    torch.cuda.empty_cache()
    
    # Return the PIL Image directly - ChatInterface will handle it properly
    return image

# --- UI Definition using gr.ChatInterface ---

seed_slider = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=42)
randomize_checkbox = gr.Checkbox(label="Randomize seed", value=False)
guidance_slider = gr.Slider(label="Guidance Scale", minimum=1.0, maximum=10.0, step=0.1, value=2.5)
steps_slider = gr.Slider(label="Steps", minimum=1, maximum=30, value=28, step=1)

# --- Examples without external URLs ---
# Remove examples temporarily to avoid format issues
examples = None

demo = gr.ChatInterface(
    fn=chat_fn,
    title="FLUX.1 Kontext [dev]",
    description="""<p style='text-align: center;'>
    A simple chat UI for the <b>FLUX.1 Kontext</b> model running on ZeroGPU.
    <br>
    To edit an image, upload it and type your instructions (e.g., "Add a hat").
    <br>
    To generate an image, just type a prompt (e.g., "A photo of an astronaut on a horse").
    <br>
    Find the model on <a href='https://huggingface.co/black-forest-labs/FLUX.1-Kontext-dev' target='_blank'>Hugging Face</a>.
    </p>""",
    multimodal=True,  # This is important for MultimodalTextbox to work
    textbox=gr.MultimodalTextbox(
        file_types=["image"],
        placeholder="Type a prompt and/or upload an image...",
        render=False
    ),
    additional_inputs=[
        seed_slider,
        randomize_checkbox,
        guidance_slider,
        steps_slider
    ],
    examples=examples,
    theme="soft"
)

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