VGGT-MPS

This VGGT-MPS server provides 3D vision and reconstruction capabilities optimized for Apple Silicon through MPS acceleration, implementing the VGGT (Vision-based Geometry and Tracking Transformer) model for multi-view 3D reconstruction from image sequences. Built with PyTorch and FastMCP, it offers tools for camera pose estimation, depth prediction, 3D point cloud generation, and point tracking across video frames, featuring specialized MPS optimizations for efficient inference on Apple's Metal Performance Shaders framework. The implementation includes Gradio web interfaces, COLMAP integration for structure-from-motion workflows, and sparse attention mechanisms for scaling to large image collections, making it valuable for researchers and developers working on 3D computer vision applications who need GPU-accelerated reconstruction on Mac hardware without CUDA dependencies.