← Back to Computer Vision
cs.CV

How to keep 3D reconstructions consistent as video streams flow

Kaichen Zhou, Zeyang Bai, Xinhai Chang, Mengyu Wang, Paul Liang, Fangneng Zhan

May 20, 2026

Video streams break existing 3D object generators—applying them frame-by-frame produces flickering, inconsistent results. Stream3D fixes this by maintaining a compact memory of the most informative historical frames, pruning outdated ones to keep memory constant as the stream lengthens. The system works with any frozen generator (SAM 3D, TRELLIS, Hunyuan3D) and outperforms simpler baselines on both real and synthetic video benchmarks without modifying the underlying model.
Published as Stream3D: Sequential Multi-View 3D Generation via Evidential Memory arXiv:2605.21472
Read the original paper →