Tencent ARC paper cuts video model training compute 57.9% while boosting physics plausibility 11.2%
_akhaliq · x · 2026-10-22
The paper "Looking Ahead: Modeling Future Evolution in Autoregressive Video Generation" teaches autoregressive video models to anticipate future dynamics via generative future prediction. It improves physical plausibility by +11.2% (I2V) and +9.4% (V2V) on Physics-IQ, while cutting training compute by 57.9% and data by 63.4%. Open-sourced by TencentARC with code and project page available; authors note current training data remains mostly general-purpose scenarios.
More from Multimodal
- LiDAR scan + LLM workflow produces dimensionally accurate 3D room models — OpenAIDevs · 2026-10-22
- MERZraum: A Poetic 3D Project Built With AI Gaussian Splatting — Merzmensch · 2026-10-22
- ComfyUI Tutorial: Model Downloads, Folder Structure and Loaders Explained — pixaromadesign · 2026-10-22
- Side-by-side: Seedance 2.5 output vs free Muse looks nearly identical — sven_ai · 2026-10-22
- HeyGen ships a wave of updates: Hyperframes Studio, HeyGen Video and Pro Voice — toolstelegraph · 2026-10-22
- Layout JSON vs plain prompt, 8 seeds each: when bounding boxes are worth it — SquareBat780 · 2026-10-22