Description
Recurrent models that reconstruct scenes frame by frame forget and drift on long sequences. TTT3R treats 3D reconstruction as test-time training, using a simple state update rule to improve CUT3R's length generalization.
No retraining is needed; it runs on CUT3R checkpoints and was published at ICLR 2026.
Long sequences:Less drift.
Training-free:Reuses CUT3R weights.
Demo:One command.
No retraining is needed; it runs on CUT3R checkpoints and was published at ICLR 2026.
Features
Long sequences:Less drift.
Training-free:Reuses CUT3R weights.
Demo:One command.
