Ring Forcing: Towards Precise Long-Term Memory for Autoregressive Video Diffusion
Ring Forcing targets the memory failure that makes long generated videos lose object identity.
The paper frames the problem as two linked gaps: object permanence when something leaves and returns, and enough memory capacity to use distant context. Its ring-structured training forces the model to retrieve from far back in the video history. The authors add compression, timestep composition, and sparse RoPE to stretch usable context to minutes-long generations under fixed sequence limits. They report stronger minutes-long coherence and object permanence than state-of-the-art methods. HF Daily Papers' note
The paper frames the problem as two linked gaps: object permanence when something leaves and returns, and enough memory capacity to use distant context. Its ring-structured training forces the model to retrieve from far back in the video history. The authors add compression, timestep composition, and sparse RoPE to stretch usable context to minutes-long generations under fixed sequence limits. They report stronger minutes-long coherence and object permanence than state-of-the-art methods. HF Daily Papers' note
score 5