Stream3D: Sequential Multi-View 3D Generation via Evidential Memory

Researchers introduced Stream3D, a training-free method that turns frozen view-conditioned 3D generators into streaming systems for long monocular video.

Stream3D caches the most informative past frames in a fixed-size evidential memory, helping preserve temporal, visual, and geometric consistency across chunks.

On synthetic and real benchmarks, it outperformed baseline memory-reuse and feature-editing methods for long-sequence 3D reconstruction.

The approach improves streaming 3D generation without retraining or modifying the base model, while keeping memory growth under control.