LayerRecall: A State-Conditioned Memory Router for Long-Horizon Consistency in Video Generation
LayerRecall routes old video memory only into the model layers that appear to need it.
The paper targets a failure mode in autoregressive video diffusion: recent-context caches keep motion smooth, but drop older cues when a subject, object, scene, or attribute returns. LayerRecall retrieves historical K/V states based on the current state, then injects them selectively into memory-sensitive DiT layers while leaving local attention intact elsewhere. The authors train the router with Cross-Horizon Prediction Matching, using a long-context reference instead of explicit memory labels. They report top overall results on MemoBench and MovieBench across 100 multi-shot prompts, with negligible inference overhead and no loss versus the backbone on VBench-Long. HF Daily Papers' note
The paper targets a failure mode in autoregressive video diffusion: recent-context caches keep motion smooth, but drop older cues when a subject, object, scene, or attribute returns. LayerRecall retrieves historical K/V states based on the current state, then injects them selectively into memory-sensitive DiT layers while leaving local attention intact elsewhere. The authors train the router with Cross-Horizon Prediction Matching, using a long-context reference instead of explicit memory labels. They report top overall results on MemoBench and MovieBench across 100 multi-shot prompts, with negligible inference overhead and no loss versus the backbone on VBench-Long. HF Daily Papers' note
score 5