World-T³ adds persistent pose-keyed memory to video generation, lifting revisit PSNR from 13.52 to 18.31 dB
jiqizhixin · x · 2026-11-12
A preprint from Meta Reality Labs, UIC and Brown studies persistent memory in camera-controlled video generation with World-T³. When a video block leaves the active context, the model writes pose-keyed scene features into fast weights via recursive least squares, retrieving them later by camera pose; the backbone stays frozen while only memory evolves, with dual-teacher distillation balancing short-range quality and long-range consistency.
- Revisit PSNR rises from SANA-WM's 13.52 to 18.31 dB on a 120-clip out-and-back camera benchmark
- The four-step model hits 20.6 FPS at 1280×704 on a single NVIDIA B200
- Memory size and per-step read/write costs stay fixed regardless of video length, though the authors note this doesn't imply flawless unlimited recall
More from Research
- After 17 years in bio, a scaling researcher calls AI-biomed hype Theranos-level — zacharynado · 2026-11-12
- AI 2027 authors: intelligence explosion on track for 2027, takeover window 3-12 months — Puzzleheaded-King584 · 2026-11-12
- sokrypton: solving biology needs subatomic simulators or massive wet labs for AI — sokrypton · 2026-11-12
- Uncertainty-Aware RWM: taming world-model hallucination for real robot control — breadli428 · 2026-11-12
- Can You Guess All Four Gradient Distributions? A Viral ML Fundamentals Quiz — cneuralnetwork · 2026-11-12
- Three neurons control Gemma 4's caring vs contempt, exported as <70KB llama.cpp vectors — New_Alps_5655 · 2026-11-12