CUT3R

CUT3R (Continuous 3D Perception Model with Persistent State) was introduced by researchers at UC Berkeley and Google DeepMind in the January 21, 2025 arXiv paper (2501.12387), later presented as a CVPR 2025 oral. It is an online 3D perception framework: a recurrent model processes a stream of images, reconstructs 3D scenes from few observations, continuously refines the reconstruction as more views arrive, and can even infer structure for unobserved regions via virtual camera queries.

The approach unifies feed-forward reconstruction with a persistent state mechanism, handling dynamic objects, sparse observations, and degenerate camera motion that challenge traditional pipelines. The authors released code on GitHub and checkpoints via Google Drive, with two published checkpoints (a 224px linear intermediate and a 512px DPT final). Check the repo for the exact license before reuse.