app.py

import gradio as gr
import torch
from transformers import AutoModel, AutoTokenizer
import cv2
import numpy as np
import tempfile
import os
import json
import time
from datetime import datetime
import ffmpeg
import soundfile as sf
from PIL import Image
import requests
import base64
import io

# Initialize MiniCPM-o model
def load_model():
    try:
        # Load MiniCPM-o 2.6 model
        model_name = "openbmb/MiniCPM-o-2_6"
        
        model = AutoModel.from_pretrained(
            model_name,
            trust_remote_code=True,
            torch_dtype=torch.float16,
            device_map="auto",
            low_cpu_mem_usage=True
        )
        
        tokenizer = AutoTokenizer.from_pretrained(
            model_name,
            trust_remote_code=True
        )
        
        return model, tokenizer
    except Exception as e:
        print(f"Error loading model: {e}")
        return None, None

# Global model loading
print("Loading MiniCPM-o 2.6 model...")
model, tokenizer = load_model()
print("Model loaded successfully!" if model else "Failed to load model")

def extract_frames_from_video(video_path, max_frames=30):
    """Extract frames from video at 1fps"""
    frames = []
    timestamps = []
    
    try:
        cap = cv2.VideoCapture(video_path)
        fps = cap.get(cv2.CAP_PROP_FPS)
        frame_interval = int(fps)  # Extract 1 frame per second
        
        frame_count = 0
        extracted_count = 0
        
        while cap.isOpened() and extracted_count < max_frames:
            ret, frame = cap.read()
            if not ret:
                break
            
            if frame_count % frame_interval == 0:
                # Convert BGR to RGB
                frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
                frames.append(Image.fromarray(frame_rgb))
                timestamps.append(extracted_count)
                extracted_count += 1
            
            frame_count += 1
        
        cap.release()
        return frames, timestamps
    except Exception as e:
        print(f"Error extracting frames: {e}")
        return [], []

def extract_audio_from_video(video_path):
    """Extract audio from video"""
    try:
        audio_path = video_path.replace('.mp4', '_audio.wav')
        
        # Use ffmpeg to extract audio
        stream = ffmpeg.input(video_path)
        stream = ffmpeg.output(stream, audio_path, acodec='pcm_s16le', ac=1, ar='16000')
        ffmpeg.run(stream, overwrite_output=True, quiet=True)
        
        return audio_path
    except Exception as e:
        print(f"Error extracting audio: {e}")
        return None

def analyze_multimodal_content(frames, timestamps, audio_path=None):
    """Analyze video frames and audio using MiniCPM-o"""
    if not model or not tokenizer:
        return "Model not loaded. Please check the model initialization."
    
    try:
        analysis_results = []
        
        # Prepare multimodal input
        for i, (frame, timestamp) in enumerate(zip(frames, timestamps)):
            # Create analysis prompt
            prompt = f"""You are an expert video narrative analyst specializing in marketing video analysis. 
            
            Analyze this frame (timestamp: {timestamp}s) and provide:
            
            🎬 NARRATIVE ANALYSIS:
            - What story moment is happening?
            - What narrative function does this serve?
            - How does this fit in the overall marketing flow?
            
            🎨 VISUAL PSYCHOLOGY:
            - What specific visual techniques are used?
            - How do colors, composition, and lighting affect emotions?
            - What psychological triggers are present?
            
            🔗 MARKETING MECHANICS:
            - How does this frame contribute to persuasion?
            - What call-to-action elements are present?
            - How does this build toward conversion?
            
            Be specific and actionable in your analysis."""
            
            try:
                # If audio is available, include it in the analysis
                if audio_path and i == 0:  # Include audio context for first frame
                    # For now, we'll do image-only analysis
                    # Future enhancement: include audio analysis
                    pass
                
                # Prepare messages for the model
                msgs = [{'role': 'user', 'content': prompt}]
                
                # Generate analysis
                response = model.chat(
                    image=frame,
                    msgs=msgs,
                    tokenizer=tokenizer,
                    sampling=True,
                    temperature=0.7,
                    max_new_tokens=500
                )
                
                analysis_results.append({
                    'frame': i + 1,
                    'timestamp': f"{timestamp}s",
                    'analysis': response[0] if isinstance(response, tuple) else response
                })
                
            except Exception as e:
                print(f"Error analyzing frame {i}: {e}")
                analysis_results.append({
                    'frame': i + 1,
                    'timestamp': f"{timestamp}s",
                    'analysis': f"Error analyzing frame: {str(e)}"
                })
        
        return analysis_results
    
    except Exception as e:
        return f"Error in multimodal analysis: {str(e)}"

def generate_comprehensive_summary(analysis_results):
    """Generate comprehensive summary using MiniCPM-o"""
    if not model or not tokenizer:
        return "Model not loaded for summary generation."
    
    try:
        # Combine all frame analyses
        combined_analysis = "\n\n".join([
            f"Frame {result['frame']} ({result['timestamp']}): {result['analysis']}"
            for result in analysis_results
        ])
        
        summary_prompt = f"""Based on the detailed frame-by-frame analysis below, provide a comprehensive marketing video analysis:

        📖 STORY ARCHITECTURE:
        - What is the overall narrative structure?
        - How does the story progress from beginning to end?
        - What transformation or journey is presented?

        🎯 PERSUASION STRATEGY:
        - What psychological principles are used?
        - How does the video build toward conversion?
        - What specific persuasion techniques are employed?

        🎨 VISUAL STORYTELLING:
        - How do visual elements support the narrative?
        - What cinematic techniques enhance the message?
        - How does the visual flow create emotional impact?

        🚀 MARKETING EFFECTIVENESS:
        - What makes this video compelling?
        - How does it capture and maintain attention?
        - What specific elements drive viewer action?

        Frame Analysis:
        {combined_analysis}

        Provide specific, actionable insights in 300 words or less."""
        
        msgs = [{'role': 'user', 'content': summary_prompt}]
        
        response = model.chat(
            image=None,
            msgs=msgs,
            tokenizer=tokenizer,
            sampling=True,
            temperature=0.3,
            max_new_tokens=600
        )
        
        return response[0] if isinstance(response, tuple) else response
    
    except Exception as e:
        return f"Error generating summary: {str(e)}"

def process_video_with_minicpm(video_file):
    """Main processing function for video analysis"""
    if video_file is None:
        return "Please upload a video file.", "", ""
    
    try:
        start_time = time.time()
        
        # Extract frames
        update_status = "Extracting frames from video..."
        frames, timestamps = extract_frames_from_video(video_file.name)
        
        if not frames:
            return "Failed to extract frames from video.", "", ""
        
        # Extract audio
        update_status = "Extracting audio from video..."
        audio_path = extract_audio_from_video(video_file.name)
        
        # Analyze with MiniCPM-o
        update_status = "Analyzing content with MiniCPM-o..."
        analysis_results = analyze_multimodal_content(frames, timestamps, audio_path)
        
        if isinstance(analysis_results, str):  # Error case
            return analysis_results, "", ""
        
        # Generate comprehensive summary
        update_status = "Generating comprehensive summary..."
        comprehensive_summary = generate_comprehensive_summary(analysis_results)
        
        # Format frame-by-frame results
        frame_analysis = "\n\n".join([
            f"🎬 **Frame {result['frame']} ({result['timestamp']})**\n{result['analysis']}"
            for result in analysis_results
        ])
        
        processing_time = time.time() - start_time
        
        # Create final report
        final_report = f"""
# 🎬 MiniCPM-o Video Analysis Report

**Analysis completed in {processing_time:.1f} seconds**
**Frames analyzed: {len(frames)}**
**Model: MiniCPM-o 2.6**

## 📊 Comprehensive Summary

{comprehensive_summary}

---

## 🎯 Technical Details

- **Processing Time**: {processing_time:.1f} seconds
- **Frames Extracted**: {len(frames)}
- **Audio Extracted**: {"Yes" if audio_path else "No"}
- **Model Used**: MiniCPM-o 2.6 (Multimodal)
- **Analysis Type**: Hybrid Audio-Visual

---

*Analysis powered by MiniCPM-o 2.6 - A GPT-4o Level MLLM*
        """
        
        return final_report, frame_analysis, comprehensive_summary
    
    except Exception as e:
        return f"Error processing video: {str(e)}", "", ""

# Create Gradio interface
def create_interface():
    with gr.Blocks(title="MiniCPM-o Video Analyzer", theme=gr.themes.Soft()) as demo:
        gr.Markdown("""
        # 🎬 MiniCPM-o Video Analyzer
        
        **Test MiniCPM-o 2.6 for advanced video analysis**
        
        Upload a marketing video (up to 30 seconds) to get:
        - 🎯 Frame-by-frame narrative analysis
        - 🎨 Visual psychology insights  
        - 🚀 Marketing effectiveness analysis
        - 📊 Comprehensive summary
        
        *Powered by MiniCPM-o 2.6 - Local multimodal analysis*
        """)
        
        with gr.Row():
            with gr.Column(scale=1):
                video_input = gr.Video(
                    label="Upload Marketing Video",
                    sources=["upload"],
                    include_audio=True
                )
                
                analyze_btn = gr.Button(
                    "🚀 Analyze with MiniCPM-o", 
                    variant="primary",
                    size="lg"
                )
                
                gr.Markdown("""
                **Tips:**
                - Upload videos up to 30 seconds for optimal analysis
                - MP4 format recommended
                - Include audio for comprehensive analysis
                """)
            
            with gr.Column(scale=2):
                with gr.Tabs():
                    with gr.TabItem("📊 Analysis Report"):
                        report_output = gr.Markdown(
                            label="Comprehensive Analysis Report",
                            value="Upload a video and click 'Analyze with MiniCPM-o' to get started."
                        )
                    
                    with gr.TabItem("🎬 Frame Analysis"):
                        frame_output = gr.Markdown(
                            label="Frame-by-Frame Analysis",
                            value="Detailed analysis of each frame will appear here."
                        )
                    
                    with gr.TabItem("📝 Summary"):
                        summary_output = gr.Markdown(
                            label="Executive Summary",
                            value="Marketing effectiveness summary will appear here."
                        )
        
        # Event handlers
        analyze_btn.click(
            fn=process_video_with_minicpm,
            inputs=[video_input],
            outputs=[report_output, frame_output, summary_output]
        )
        
        # Examples
        gr.Markdown("""
        ## 🎯 What This Analysis Provides
        
        - **Narrative Analysis**: Story structure and progression
        - **Visual Psychology**: Color, composition, and emotional triggers
        - **Marketing Mechanics**: Persuasion techniques and conversion strategies
        - **Attention Engineering**: How the video captures and maintains viewer focus
        - **Comparative Insights**: How this compares to your existing GPT-4o analysis
        """)
    
    return demo

# Launch the app
if __name__ == "__main__":
    demo = create_interface()
    demo.launch(
        server_name="0.0.0.0",
        server_port=7860,
        share=True
    )