docker-ai-models

Docker containers for AI models with GPU support.

This repository provides pre-configured Docker containers for various AI models. Each model includes:

• Complete environment setup with all dependencies
• GPU acceleration support via CUDA
• Web-based interface for interaction
• Persistent model storage to avoid re-downloading

The containers handle dependency management, CUDA configuration, and model downloads automatically. Models are stored in mounted volumes for reuse across container rebuilds.

Target hardware: Consumer GPUs (RTX series). All containers tested on RTX 3060 12GB VRAM.

Memory optimization: Uses mmgp (Memory Management for GPU Poor) to run large models on consumer hardware. mmgp enables models that normally require 24GB+ VRAM to run on 12GB cards through:

• 8-bit quantization of large model components
• Dynamic model offloading between VRAM and system RAM
• Selective layer loading (loads only active layers to VRAM)
• Reserved RAM pinning for fast transfers

Example: FLUX.1-schnell (22.7GB transformer + 8.8GB text encoder) runs on 12GB VRAM via quantization and partial pinning to system RAM.

Use case: Run inference on large AI models locally using retail GPUs without requiring datacenter hardware.

---

Repository Structure

Each model directory contains:

• dockerfile - Container build configuration
• startup.sh - Container initialization script
• gradio_interface.py - Web UI implementation
• requirements.txt - Python dependencies

Directory layout:

├── yue/                    # YuE: Lyrics-to-song music generation
│   ├── dockerfile          # CUDA 12.4 runtime (12GB VRAM)
│   ├── gradio_interface.py # Web interface + generation pipeline
│   ├── startup.sh          # Container entrypoint
│   └── requirements.txt    # Python dependencies
├── seed-story/             # SEED-Story: Comic story generation
│   ├── dockerfile          # GPU build (8GB+ VRAM)
│   ├── minimal_gradio.py   # Main web interface
│   ├── simple_comic_generator.py # Fallback generator
│   ├── model_downloader.py # Model management
│   ├── startup.sh          # Launch script
│   └── requirements.txt    # Python dependencies
├── omnicontrol/            # OmniControlGP: Subject-driven image generation
│   ├── dockerfile          # GPU build (12GB+ VRAM)
│   ├── gradio_interface.py # Web interface
│   ├── startup.sh          # Launch script
│   ├── src/flux/           # FLUX model source code
│   └── requirements.txt    # Python dependencies
├── wan/                    # Wan2GP: Video-to-audio generation
│   ├── dockerfile-gpu-poor # GPU-efficient build
│   └── startup-gpu-poor.sh # Launch script

---

Available Models

Model	Type	Hardware	Interface	Documentation
🎵 YuE	Lyrics-to-Song Generation	12GB VRAM	Web UI	Setup Guide
🎬 SEED-Story	Comic Story Generation	8GB+ VRAM	Web UI	Setup Guide
🎨 OmniControl	Subject-Driven Image Gen	12GB+ VRAM	Web UI	Setup Guide
🎵️ Wan2GP	Video-to-Audio Synthesis	8GB+ VRAM	Web UI	Coming Soon

---

Planned Models

Model	Type	Description	Repository
📸 PhotoMaker	Image Generation	Customizing realistic human photos via stochastic identity mixing	TencentARC/PhotoMaker
🗣️ Fantasy Talking	Video Generation	High-quality talking face generation with identity preservation	Fantasy-AMAP/fantasy-talking

---

Quick Start

# Clone the repository
git clone https://github.com/Ricky-G/docker-ai-models.git
cd docker-ai-models

# Navigate to your chosen model
cd omnicontrol/  # or yue/ or seed-story/ or wan/

# Build and run (see individual README files for specific commands)
docker build -t omnicontrol .
docker run -d --gpus all -p 7860:7860 \
  -v D:\_Models\omnicontrol:/app/models \
  -e HF_TOKEN=your_token_here \
  omnicontrol

Web interface available at http://localhost:7860

---

Requirements

• Docker with NVIDIA GPU support
• NVIDIA Container Toolkit
• GPU with sufficient VRAM (see model-specific requirements)

---

Contributing

To add a new model:

1. Fork the repository
2. Create a new directory for your model
3. Add dockerfile, startup script, and README
4. Submit a pull request

---

License

MIT License