LayerRecall: A State-Conditioned Memory Router for Long-Horizon Consistency in Video Generation

TL;DR - LayerRecall is a state-conditioned memory router that helps autoregressive video diffusion models recover relevant distant context by injecting retrieved historical K/V states into selected transformer layers. It improves long-horizon consistency with negligible inference overhead while preserving local motion and scene continuity.

• Identifies distinct layer-level preferences for current, recent, and distant context in video diffusion transformers.
• Routes relevant historical states only to memory-sensitive layers, leaving local attention unchanged elsewhere.
• Uses Cross-Horizon Prediction Matching to train from a privileged long-context model without explicit memory-allocation labels.
• Achieves the best reported overall results on MemoBench and MovieBench across 100 multi-shot prompts while matching its backbone on VBench-Long.