MosaiChunk: compositing spatio-temporal KV memory for autoregressive video generation
camenduru · x · 2026-10-13
Researchers from Cornell, UC Berkeley, Harvard and Impossible (including Alexei Efros and Qianqian Wang) release MosaiChunk, tackling detail loss in long-horizon autoregressive video generation when objects leave the context window.
- Idea: instead of retrieving whole past chunks, compose a mosaic of selected historical key–value entries across space and time; a frozen video generator consumes these non-contiguous KV entries directly to recover visual details.
- Method: keep the generator frozen and train only a lightweight router to select memory under a fixed active-memory budget.
- Demo: e.g. a star cookie keeps its appearance when a tin closes and reopens.
Webpage, paper, code, weights and data are all released openly.
Related event: MosaiChunk Gives Autoregressive Video Models Persistent Visual Memory(2 posts)→
More from Multimodal
- Meta's Muse Wonder (Max) tested on one-shot 3D scenes and front-end vs top models — arena · 2026-10-13
- AI video recreates Columbus's 1492 landing proclamation from his actual journal — aeeiee · 2026-10-13
- Open-source VAE fixes Qwen-Image 2.1's checkerboard artifacts — barbear22 · 2026-10-13
- ComfyUI + MiniMax H3 + Claude: A Short-Form Video Pipeline Demo — TheHollywoodGeek · 2026-10-13
- SpatialASM: Shanghai AI Lab trains LVLMs to actively mentalize viewpoints for spatial reasoning — Shanghai-AI-Laboratory · 2026-10-13
- Paper turned into ComfyUI plugin: more diverse images across diffusion models at no extra cost — michaelchchoi · 2026-10-13