Video Caption Suite

Batch video captioning using Qwen3-VL-8B vision-language model. Select a directory, process videos, get captions saved alongside them.

Requirements

• Python 3.10+
• CUDA-capable GPU (8GB+ VRAM recommended)
• Node.js 18+ (for frontend build)

Installation

Windows:

install.bat

Linux/Mac:

chmod +x install.sh
./install.sh

This creates a virtual environment and installs all dependencies.

Usage

Windows:

start.bat

Linux/Mac:

./start.sh

Open http://localhost:8000 in your browser.

Workflow

1. Click Settings and select your working directory
2. Videos from that directory appear in the grid
3. Select videos and click Process
4. Captions are saved as .txt files alongside the videos

Configuration

Edit config.py to adjust:

Setting	Default	Description
MODEL_ID	Qwen/Qwen3-VL-8B-Instruct	HuggingFace model
MAX_FRAMES_PER_VIDEO	128	Frames extracted per video
FRAME_SIZE	336	Frame dimension in pixels
MAX_TOKENS	512	Max caption length
TEMPERATURE	0.3	Generation creativity

Multi-GPU Processing

On systems with multiple CUDA GPUs, the suite automatically detects available devices and enables parallel processing:

• Auto-detection: GPUs are detected on startup via /api/system/gpu
• Batch size: Set how many videos to process simultaneously (1 per GPU, max 8)
• Parallel workers: Each GPU loads its own model copy and processes videos independently

The batch size slider appears in Settings → Optimization only when multiple GPUs are detected. Each GPU requires ~16GB VRAM to hold the Qwen3-VL-8B model.

Project Structure

Video Caption Suite/
├── backend/          # FastAPI server
├── frontend/         # Vue 3 UI
├── models/           # Downloaded model cache
├── config.py         # Settings
├── install.bat/sh    # Installation
└── start.bat/sh      # Launch server

License

MIT