app.py

import os
import torch
import gradio as gr
import spaces  # Critical for ZeroGPU support in Spaces
from diffusers import DiffusionPipeline  # Common for HF video models
from PIL import Image

# =============================================================
# INITIALIZATION & PIPELINE CONFIGURATION
# =============================================================
MODEL_ID = "MiniMaxAI/MiniMax-M1"

def load_pipeline():
    # Load model with bfloat16 for modern GPU efficiency [17]
    # Note: Ensure the model is available on the HF Hub or adjust path
    pipe = DiffusionPipeline.from_pretrained(
        MODEL_ID, 
        torch_dtype=torch.bfloat16,
        use_safetensors=True
    )
    
    # Critical VRAM optimizations for deployment on 24GB or ZeroGPU [2]
    pipe.enable_model_cpu_offload() 
    pipe.enable_vae_slicing()
    return pipe

# Pipeline is initialized globally to avoid reloads on every click
pipe = load_pipeline()

# =============================================================
# GENERATION LOGIC
# =============================================================
@spaces.GPU(duration=300)  # Allocated GPU time for complex generation [16]
def generate_video(prompt, negative_prompt, steps, guidance_scale, seed):
    # Standard 5-second video length at 16 FPS [18-20]
    num_frames = 81 
    
    # Use Generator for deterministic results [21, 22]
    generator = torch.Generator("cuda").manual_seed(int(seed))
    
    # Execution using VACE-style inputs [8, 23]
    output = pipe(
        prompt=prompt,
        negative_prompt=negative_prompt,
        num_inference_steps=int(steps),
        guidance_scale=float(guidance_scale),
        num_frames=num_frames,
        generator=generator
    ).frames
    
    # Exporting generated frames to a video file [24, 25]
    # For a real app, use diffusers.utils.export_to_video
    from diffusers.utils import export_to_video
    import tempfile
    
    temp_path = tempfile.mktemp(suffix=".mp4")
    export_to_video(output, temp_path, fps=16)
    return temp_path

# =============================================================
# UI DESIGN (Using gr.Blocks for Professional Layout)
# =============================================================
# CSS for custom styling to improve "Joyful Experience" [26, 27]
css = """
.container { max-width: 1000px; margin: auto; }
.gen_btn { background-color: #7224f2 !important; color: white !important; }
"""

with gr.Blocks(css=css, theme=gr.themes.Soft()) as demo:
    gr.Markdown(f"# {MODEL_ID} High-Fidelity Video Generator")
    gr.Markdown("Leveraging ZeroGPU and VRAM offloading for cinematic AI video [2, 28].")
    
    with gr.Row():
        with gr.Column(scale=1):
            prompt_input = gr.Textbox(
                label="Prompt", 
                placeholder="Describe the action and scene details...",
                lines=4
            )
            neg_prompt = gr.Textbox(
                label="Negative Prompt", 
                value="blurry, distorted, low quality, watermark, text"
            )
            
            with gr.Accordion("Advanced Settings", open=False):
                steps = gr.Slider(20, 50, value=30, step=1, label="Inference Steps")
                guidance = gr.Slider(1.0, 15.0, value=7.0, label="Guidance Scale")
                seed = gr.Number(value=42, label="Seed")
            
            generate_btn = gr.Button("Generate Video", variant="primary", elem_classes="gen_btn")
            
        with gr.Column(scale=1):
            video_output = gr.Video(label="Generated Output")

    # Event listener mapping [29, 30]
    generate_btn.click(
        fn=generate_video,
        inputs=[prompt_input, neg_prompt, steps, guidance, seed],
        outputs=video_output
    )

# =============================================================
# LAUNCH
# =============================================================
if __name__ == "__main__":
    # Ensure app.py is at the root for automatic HF detection [14, 15]
    demo.launch()