app.py

import io
from pathlib import Path
from typing import Tuple, Optional
import gradio as gr
import numpy as np
import torch
from PIL import Image
from dotenv import load_dotenv
from diffusers import DiffusionPipeline
from transformers import pipeline
from huggingface_hub import login
import os

# Load environment variables
load_dotenv()
hf_token = os.getenv("HF_TKN")
if hf_token:
    login(token=hf_token)

# Device configuration
device = "cuda" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

# Load models
@spaces.GPU
def load_models():
    """Load both models with proper device placement"""
    caption_pipe = pipeline(
        "image-to-text",
        model="nlpconnect/vit-gpt2-image-captioning",
        device=device
    )

    audio_pipe = DiffusionPipeline.from_pretrained(
        "cvssp/audioldm2",
        token=hf_token,
        torch_dtype=torch_dtype
    )
    return caption_pipe, audio_pipe

caption_pipe, audio_pipe = load_models()

def analyze_image(image_bytes: bytes) -> Tuple[str, bool]:
    """Generate caption from image bytes with enhanced error handling"""
    try:
        image = Image.open(io.BytesIO(image_bytes))
        if image.mode != "RGB":
            image = image.convert("RGB")
            
        results = caption_pipe(image)
        
        if not results or not isinstance(results, list):
            return "Error: Invalid response from caption model", True
            
        caption = results[0].get("generated_text", "").strip()
        return caption or "No caption generated", not bool(caption)
        
    except Exception as e:
        return f"Image processing error: {str(e)}", True

@spaces.GPU(duration=120)
def generate_audio(caption: str) -> Optional[Tuple[int, np.ndarray]]:
    """Generate audio from caption with resource management"""
    try:
        # Device management with context
        original_device = next(audio_pipe.parameters()).device
        audio_pipe.to(device)
        
        # Generation with progress awareness
        audio = audio_pipe(
            prompt=caption,
            num_inference_steps=50,
            guidance_scale=7.5,
            audio_length_in_s=5.0  # Keep audio generation short
        ).audios[0]
        
        # Post-processing
        audio = audio.squeeze()  # Handle mono channel
        audio = np.clip(audio, -1, 1)  # Ensure valid range
        return (16000, audio)
        
    except Exception as e:
        print(f"Audio generation error: {str(e)}")
        return None
        
    finally:
        audio_pipe.to(original_device)
        if torch.cuda.is_available():
            torch.cuda.empty_cache()

# UI Components
css = """
#col-container {
    max-width: 800px; 
    margin: 0 auto;
}
.disclaimer {
    font-size: 0.9em; 
    color: #666;
}
"""

with gr.Blocks(css=css) as demo:
    with gr.Column(elem_id="col-container"):
        gr.HTML("""
            <h1 style="text-align: center;">🎶 Image to Sound Effect Generator</h1>
            <p style="text-align: center;">
                ⚡ Powered by <a href="https://bilsimaging.com" target="_blank">Bilsimaging</a>
            </p>
        """)
        
        with gr.Row():
            image_input = gr.Image(type="filepath", label="Upload Image")
            caption_output = gr.Textbox(label="Generated Description", interactive=False)
        
        with gr.Row():
            generate_btn = gr.Button("Generate Description", variant="primary")
            audio_output = gr.Audio(label="Generated Sound", interactive=False)
            sound_btn = gr.Button("Generate Sound", variant="secondary")
        
        gr.Examples(
            examples=[str(Path(__file__).parent / "examples" / f) for f in ["storm.jpg", "city.jpg"]],
            inputs=image_input,
            outputs=[caption_output, audio_output],
            fn=lambda x: (analyze_image(Path(x).read_bytes())[0], None),
            cache_examples=True
        )

        gr.Markdown("### 🛠️ Usage Tips")
        gr.Markdown("""
            - Use clear, high-contrast images for best results
            - Complex scenes may require multiple generations
            - Keep sound generation under 10 seconds for quick results
        """)

        gr.Markdown("### ⚠️ Disclaimer", elem_classes="disclaimer")
        gr.Markdown("""
            Generated content may not always be accurate. Use at your own discretion.
            [Privacy Policy](https://bilsimaging.com/privacy) | 
            [Terms of Service](https://bilsimaging.com/terms)
        """)

    # Event handling
    generate_btn.click(
        fn=lambda x: analyze_image(Path(x).read_bytes())[0],
        inputs=image_input,
        outputs=caption_output,
        api_name="describe"
    )

    sound_btn.click(
        fn=generate_audio,
        inputs=caption_output,
        outputs=audio_output,
        api_name="generate_sound"
    )

    # Input validation
    image_input.change(
        fn=lambda: [gr.update(value=""), gr.update(value=None)],
        outputs=[caption_output, audio_output]
    )

if __name__ == "__main__":
    demo.launch(server_name="0.0.0.0" if os.getenv("SPACE_ID") else "127.0.0.1")