History Forcing improves long-horizon video world models
Original title: History Forcing: Distilling History-Dependent Continuations for Long-Horizon Video World Models
History Forcing is a distillation framework that strengthens historical supervision for few-step autoregressive video generators, achieving state-of-the-art results.
Few-step autoregressive video diffusion struggles with consistency in long rollouts when subjects reappear. History Forcing uses residuals between teacher predictions with and without historical context to decouple supervision. It achieves state-of-the-art performance among 1.3B models on VBench for 15s and 60s video generation.
- State-of-the-art VBench quality score of 85.07 for 15s videos
- VBench-Long score of 79.82 for 60s generation
- Improves Evolution Quality from 2.35 to 2.70
- Enhances Object Consistency from 58.43 to 62.18
From the original article
1 min read at huggingface.co
Few-step autoregressive video diffusion enables efficient generation, yet maintaining consistency over long rollouts remains challenging, especially when subjects or scenes reappear after going out of sight. Despite progress in fidelity and rollout robustness, existing distillation methods do not fully exploit richer historical information to constrain the student's current prediction.
By Hangyi Kuang, Chengzhi Wang, Yefei He · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.