TTT3R: Enhancing 3D Reconstruction with Test-Time Training for CUT3R

Test-time training 3D reconstruction

I introduce TTT3R, a simple state update rule designed to significantly improve length generalization for the CUT3R model. By treating 3D reconstruction as a test-time training process, our approach allows the system to adapt dynamically to varying sequence lengths without retraining. This method, accepted at ICLR 2026, offers a lightweight solution to a persistent challenge in visual SLAM and 3D scene understanding.