单目视图合成工具

# Pull and run (All-in-One, ~15GB with model)
docker run -d --gpus all -p 8080:8080 --name sharp neosun/sharp:latest

# Access Web UI
open http://localhost:8080

# API Documentation
open http://localhost:8080/docs

version: '3.8'
services:
  sharp:
    image: neosun/sharp:latest
    container_name: sharp-service
    ports:
      - "8080:8080"
    environment:
      - GPU_IDLE_TIMEOUT=300
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    restart: unless-stopped

docker-compose up -d

# Upload image and generate 3D
curl -X POST http://localhost:8080/api/predict \
  -F "file=@photo.jpg" \
  -F "render_video=true"

# Response
{
  "task_id": "abc123",
  "ply_url": "/api/files/abc123.ply",
  "video_url": "/api/files/abc123.mp4"
}

# Download results
curl -O http://localhost:8080/api/files/abc123.ply
curl -O http://localhost:8080/api/files/abc123.mp4

# Check GPU status
curl http://localhost:8080/api/gpu/status
# {"device":"cuda","model_loaded":true,"gpu_memory_allocated_mb":2694}

# Release GPU memory
curl -X POST http://localhost:8080/api/gpu/offload
# {"status":"offloaded"}

{
  "mcpServers": {
    "sharp": {
      "command": "docker",
      "args": ["exec", "-i", "sharp-service", "python", "mcp_server.py"]
    }
  }
}

User: Generate a 3D scene from /path/to/image.jpg

Assistant: I'll generate a 3D Gaussian Splat from that image.
[Calls predict tool with file_path="/path/to/image.jpg"]
Result: PLY file saved to /tmp/sharp/output/image.ply

sharp/
├── app.py              # Flask API server
├── gpu_manager.py      # GPU resource management
├── mcp_server.py       # MCP interface
├── templates/          # Web UI templates
├── static/             # Frontend assets
├── src/sharp/          # Core model code
├── Dockerfile          # Container definition
└── docker-compose.yml  # Compose config