Stream4D: training-free streaming 4D reconstruction via causal KV-cache motion decoupling
zhenjun_zhao · x · 2026-10-24
Stream4D is a training-free streaming 4D reconstruction method that exploits causal KV-cache attention to encode temporal motion cues.
- Motion-decoupled masking: queries load cached keys/values to compute attention scores, aggregate temporal saliency into dynamic masks, then modify cached keys to suppress dynamic-region interference during historical aggregation
- Largely outperforms prior streaming methods on DAVIS segmentation: 43.39 J&M on DAVIS-2016 (previous streaming best 34.91), 43.06/41.32 on DAVIS-all, matching or beating several offline methods
- Qualitatively smoother, more complete masks under occlusion and fast/non-rigid motion
More from Research
- UTokyo/RIKEN lab brings CAROT cross-lingual alignment method to EMNLP, +11.2 accuracy points — Bollegala · 2026-10-24
- GPT-6 Astra solves 49 of 50 research-level physics problems, catches experts' own errors — MoonL88537 · 2026-10-24
- Rice & NVIDIA's Sensorimotor Alignment predicts robot policy success without training (r=0.717) — CSProfKGD · 2026-10-24
- RLVR Trains VLMs on Instruction Hierarchy with Just 7K Verifiable Examples — TuhinChakr · 2026-10-24
- NVIDIA ships BioNeMo Inference Runtime v0.2 with RFD4-Proteina support — AllThingsApx · 2026-10-24
- Trillion-observation finance benchmark: linear ridge probe beats 11 of 14 self-supervised encoders — burny_tech · 2026-10-24