110-Page Survey From 6 Institutions Maps Memory Mechanisms in Autoregressive Video Generation
jiqizhixin · x · 2026-10-15
HKUST, CityU, Fudan, CMU, NYU and NTU released a 110-page survey, 'The Past Frames the Future: Memory for Autoregressive Video Generation' — the first comprehensive map of memory mechanisms for long-video generation. As videos grow longer, memory becomes unavoidable: does a character who leaves and re-enters stay the same person? Does spatial layout survive camera pans? Will a moved object snap back dozens of steps later? The survey defines memory as historical state maintained across autoregressive steps and organizes the field along four axes: memory functions (causal/semantic/dynamic/spatial/identity), memory forms (implicit states like KV cache, explicit 3D world states, adaptive parametric), memory operations (write/read/update/manage/integrate), and memory-aware learning paradigms. A practical selection map for anyone building long-horizon video models.
More from Multimodal
- PixText fixes text in generated images: 83% glyph pass rate vs 37% for baselines — ATH-MaaS · 2026-10-15
- TextVerifier catches flawed Chinese glyph rendering with F1 0.844, far above OCR and VLMs — ATH-MaaS · 2026-10-15
- AI Video Creators: 'Amazing Visuals Are Solved — Unique IP and Characters Are What's Left' — AIandDesign · 2026-10-15
- Microsoft's MAI-Image-2.6 and Flash land on Vercel AI Gateway with one API key — lmoroney · 2026-10-15
- Stunning AI video made with Seedance 2.5 draws praise as "a masterpiece" — SimplyAnnisa · 2026-10-15
- Stunning ink-and-paper manta ray video is pure shaders: 4K in 90s for $0.40 — ravisparikh · 2026-10-15