Open-Sora 2.0 Video Generation Service

A simple web-based service for generating 10-20 second videos using Open-Sora 2.0, with image and text input.

Features

• 🎬 Generate high-quality videos (10-20 seconds)
• 🖼️ Image + Text conditioning for better control
• 🚀 Simple web interface
• 📥 Direct video download
• 🔓 No content restrictions

Architecture

• Backend: FastAPI (Python)
• Frontend: Vanilla HTML/CSS/JavaScript
• Model: Open-Sora 2.0 (11B model)

Prerequisites

• Python 3.10+
• CUDA-compatible GPU (H100/H800 recommended)
• 40GB+ VRAM for optimal performance
• PyTorch >= 2.4.0

Quick Start Options

Option 1: Remote GPU (Recommended) ⚡⚡⚡

NEW: Run frontend & backend locally, use remote GPU for generation!

This is the most cost-effective and flexible setup:

• ✅ No local GPU required
• ✅ Pay for GPU only when generating
• ✅ Works on Windows, Mac, Linux
• ✅ Quick 5-minute setup

# On Lambda GPU instance
./lambda_run_gpu_service.sh --api-key your-secret-key

# On your local machine
./run_local.sh --gpu-url http://lambda-ip:8001 --api-key your-secret-key

# Open browser
http://localhost:8000

Complete guide: See QUICKSTART_REMOTE.md and REMOTE_GPU_SETUP.md

Option 2: All-in-One Lambda Instance

Run everything (frontend + backend + GPU) on a single Lambda instance.

# 1. Launch instance at https://cloud.lambdalabs.com
# 2. SSH into instance
ssh ubuntu@<instance-ip>

# 3. Run automated setup
git clone <your-repo-url> ~/sora2
cd ~/sora2
./lambda_setup.sh

# 4. Start service
./lambda_run.sh

Complete guide: See LAMBDA_LABS.md

Option 3: Local Installation

For local GPU servers or other cloud providers.

Installation

1. Clone Open-Sora Repository

git clone https://github.com/hpcaitech/Open-Sora.git
cd Open-Sora
pip install -v .
pip install xformers==0.0.27.post2 --index-url https://download.pytorch.org/whl/cu121
pip install flash-attn --no-build-isolation

2. Install Service Dependencies

cd /path/to/sora2
pip install -r requirements.txt

3. Configure Open-Sora Path

Edit backend/config.py and set the OPENSORA_PATH to your Open-Sora installation directory.

Usage

Start the Service

# From the project root
python backend/main.py

The service will start on http://localhost:8000

Using the Web Interface

1. Open http://localhost:8000 in your browser
2. Upload an input image (PNG/JPG)
3. Enter your text prompt
4. Configure video settings:
   • Duration (10-20 seconds)
   • Aspect ratio (16:9, 9:16, 1:1, 2.39:1)
   • Motion intensity
5. Click "Generate Video"
6. Wait for generation (1-5 minutes depending on GPU)
7. Preview and download your video

API Endpoints

Generate Video

POST /api/generate
Content-Type: multipart/form-data

Parameters:
- image: File (required)
- prompt: string (required)
- duration: integer (10-20, default: 15)
- aspect_ratio: string (default: "16:9")
- motion_score: float (0.0-1.0, default: 0.5)

Response:
{
  "video_id": "uuid",
  "status": "processing"
}

Check Status

GET /api/status/{video_id}

Response:
{
  "status": "completed",
  "video_url": "/api/download/{video_id}"
}

Download Video

GET /api/download/{video_id}

Configuration

Edit backend/config.py:

OPENSORA_PATH = "/path/to/Open-Sora"
OUTPUT_DIR = "./outputs"
MAX_CONCURRENT_JOBS = 2
ENABLE_QUEUE = True

Performance Notes

• 256px: ~60 seconds on single H100/H800
• 768px: ~276 seconds with 8 GPUs
• Use --offload True for memory optimization on GPUs with <40GB VRAM

Deployment Options

• Lambda Labs (Recommended): See LAMBDA_LABS.md - Quick 5-minute setup
• Docker: See DEPLOYMENT.md
• Production: See DEPLOYMENT.md for systemd, nginx, SSL setup

Troubleshooting

Out of Memory

Add --offload True to the generation command in backend/generator.py

Slow Generation

• Use multiple GPUs with --nproc_per_node
• Reduce resolution or frame count
• Enable sequence parallelism

Model Download Issues

Models are downloaded automatically from HuggingFace. If you experience issues:

• Use ModelScope mirror (Chinese users)
• Download manually and update config paths

License

This service wrapper is provided as-is. Open-Sora 2.0 is licensed under Apache 2.0.

Credits

• Open-Sora by HPC-AI Tech