Fast3R is Meta AI’s official CVPR 2025 release for “Towards 3D Reconstruction of 1000+ Images in One Forward Pass.” It represents a next-generation feedforward 3D reconstruction model capable of producing dense point clouds and camera poses for hundreds to thousands of images or video frames in a single inference pass—eliminating the need for slow, iterative structure-from-motion pipelines. Built on PyTorch Lightning and extending concepts from DUSt3R and Spann3r, Fast3R unifies multi-view geometry, depth estimation, and camera registration within a single transformer-based architecture. It outputs high-quality 3D scene representations from unordered or sequential views, scaling to large datasets and varied camera intrinsics. The repository includes pretrained models, Gradio-based demos, and modular APIs for direct integration into research or production workflows.
Features
- Supports distributed multi-node training and evaluation via Slurm
- Exportable 3D point clouds and depth maps for downstream rendering or scene understanding
- Evaluation support for DTU, 7-Scenes, NeuralRGBD, Tanks & Temples, CO3D, and RealEstate10K
- Compatible with DUSt3R and Spann3r dataset preprocessing formats
- Full PyTorch and Lightning training pipelines with Hydra-based configuration
- Pretrained weights downloadable from Hugging Face
Project Samples
