MosaiChunk addresses a common problem in long video generation: details change when an object leaves the context window and later reappears. Its paper was first submitted on October 1, rather than during the roundup week.
The method keeps the video generator frozen and trains a lightweight router to select historical key-value entries across space and time. The authors also introduce RememBench to test scene revisits in text-to-video and image-to-video settings.
Code, router checkpoints and benchmark resources are linked by the authors. Usage terms depend on the chosen backbone and include non-commercial restrictions; improved consistency in reported tests is not a guarantee of perfect memory.
Sources: Research paper · Author code and usage terms