Online Neural Space Time Memory for Dynamic Novel View Synthesis

TL;DR - A research method for real-time online novel view synthesis from multi-view streaming video that decouples how often a neural "space-time memory" is updated versus applied, enabling long-horizon scene reconstruction under strict real-time constraints. It matters because it makes Test-Time Training-style memory practical for live, dynamic scenes.

• Identifies a core trade-off: persistent long-horizon memory (to recover temporarily occluded regions) vs. real-time budgets, noting standard TTT requires costly per-frame gradient updates that block real-time use and cause long-context instability.
• Key idea: decouple update and application frequencies — do periodic (not per-frame) memory updates while applying memory every frame, using cross-view attention to handle deformations between prior memory state and the current frame.
• Introduces two stabilizing mechanisms: an auxiliary Memory Loss to force persistent internalization of the scene, and a Memory Caching strategy that regularizes active weights against catastrophic drift.
• Claims real-time, state-of-the-art results on dynamic human-motion scenes and minute-scale online memorization (specific quantitative metrics/baselines not provided in the abstract).