DUSt3R

DUSt3R reframes multi-view 3D reconstruction as pointmap regression: a transformer with a ViT-L encoder and ViT-B decoder takes uncalibrated, unposed image pairs and directly predicts dense 3D point coordinates with confidence maps, from which camera intrinsics, poses, and depth all fall out as by-products. Presented at CVPR 2024 by Naver, the approach removed the need for the camera calibration that classical structure-from-motion pipelines require, and it seeded a wave of follow-up work including MASt3R and VGGT. The repository ships three pretrained checkpoints at 224 and 512 pixel resolutions with linear or DPT heads, a Gradio demo, Docker images with CUDA support, and full training code covering nine datasets such as CO3Dv2, ARKitScenes, and ScanNet++. Setup uses a conda environment with PyTorch and an optional compiled CUDA kernel for RoPE position embeddings, and inference wants an NVIDIA GPU. Code and weights are released under CC BY-NC-SA 4.0, so commercial use requires a separate arrangement with Naver; researchers use it for fast pose-free reconstruction, and it holds about 7.3k GitHub stars.