Now autoregressive video models can remember. 🧠 Introducing MosaiChunk: persistent visual memory for long-horizon autoregressive video generation🧩. Instead of remembering everything, it remembers the right things. ✍️ 1/6

We find that a frozen video generator can directly use irregularly and sparsely sampled historical KV as extra context to generate videos with high visual fidelity, without backbone fine-tuning. 2/6

Building on this observation, we design a memory router that automatically identifies and composes the right historical KV entries. Each chunk is divided into sections with learned descriptors. The latest chunk retrieves the most relevant sections, forming a MosaiChunk under a fixed active-memory budget. 3/6

We train only the descriptor encoder through self-distillation. Teacher and student share a frozen generator. The teacher sees whole historical chunks that fully cover the historical content to be redrawn; the student receives the router's MosaiChunk. Matching their denoising predictions trains the encoder without ground-truth target videos. 4/6

🧪 Evaluating long-horizon revisit consistency calls for a dedicated benchmark. We introduce RememBench: - T2V: 100 four-stage prompt schedules with an out-of-sight interval. - I2V: 150 camera-controlled scenes (50 indoor, 100 outdoor). MosaiChunk improves revisit consistency over a sliding window and whole-chunk retrieval (MoC). 5/6

🚀 Build video models that remember. See MosaiChunk in action and try it yourself: Webpage: Paper: Code: Weights: Data: Amazing team effort led by @YiwenZhangYZ with @HaochengXiUCB @MichaelTYLiu Alexei Efros @ElorHadar @QianqianWang5 @HavenFeng 6/6