Tencent proposes RMD distillation to fix error accumulation in long-horizon AR video generation
tencent · hf · 2026-10-05
Tencent introduces Rollout-Marginal Distillation (RMD), tackling error accumulation in autoregressive (AR) video diffusion over long rollouts.
Problem: Existing video-level DMD scores whole rollouts jointly, so a chunk's correction can be dragged toward artifacts in surrounding context just to preserve temporal consistency.
Method:
- RMD keeps generated history for AR prediction but scores each chunk independently against a chunk teacher, so quality correction isn't compromised by imperfect context;
- A subsequent video-level DMD pass restores temporal coherence.
Results: RMD maintains high visual quality well beyond its training horizon and outperforms video-level DMD baselines. Code and video results are public on the project page.
More from Multimodal
- Ego2Act benchmark tests 6 SOTA video models on goal-directed action execution — mohitban47 · 2026-10-05
- NihonSub: Open-Source Real-Time Anime Subtitle Engine Built on Whisper + DeepSeek — Grand_Marionberry115 · 2026-10-05
- RAT KING: A Character Consistency Test of MiniMax H3 — malcolmrey · 2026-10-05
- ElevenLabs launches The Search, a $100K contest to find the world's catchiest ad jingle — lukeharries · 2026-10-05
- Runway Teases Behind-the-Scenes of Jon Wertheim's New 60 Minutes Title Sequence — c_valenzuelab · 2026-10-05
- Donut Breakfast Ad Made With Google's Omni Video Model, Prompt Shared — michaelrabone · 2026-10-05