MosaiChunk

MosaiChunk is a research method for long-horizon autoregressive video generation. Such generators have a finite context window, so details of an object or scene that leaves the window can be lost and are hard to recover when it reappears. MosaiChunk addresses this with a memory mechanism that keeps the generator frozen and trains only a lightweight router.

The method splits each chunk's unrotated key–value (KV) cache into equal-size sections using balanced k-means, and stores them with learned descriptors in CPU memory. When generating a new chunk, sections of the latest chunk query older sections, and the top-scoring ones within a fixed budget are composed into a "MosaiChunk" that the frozen diffusion transformer reads alongside its sliding window. The descriptor encoder is trained by self-distillation against the same generator given richer whole-chunk memory, so no ground-truth target videos are needed.

The authors apply it to two backbones: RAVEN-adapted MiniMax-H3 (H3-AR) for text-to-video and LingBot-World-Infinity for image-to-video. They also introduce RememBench, a 100-sample benchmark in which objects appear, disappear, stay hidden for about five seconds and return. They report improved revisit consistency over a sliding-window baseline and whole-chunk retrieval (MoC) under matched memory budgets. The provided material does not show released code, weights or a paper link.