Honeycomb packs video world model memory into six fixed feature planes to stop drift
机器之心 · wechat · 2026-10-15
Researchers from Harvard, MIT and NUS propose Honeycomb, a video world model with HexMemory—a constant-size scene memory made of six fixed 2D feature planes (three spatial, three spatiotemporal).
- Instead of accumulating point clouds or latent features, the system repeatedly compresses new observations into the same tensors via a read-generate-write loop, so memory never grows with video length.
- A feed-forward writer processes only the newest clip, fusing it with prior memory by confidence maps; writes stay 13ms regardless of clip count, 240x faster than per-scene optimization.
- On WorldScore closed-loop revisit tests it hits 17.22dB PSNR (1.23dB above the runner-up) with 3.00px optical-flow error vs Spatia's 6.64; it also leads on RealEstate10K novel-view synthesis and scores 65.52 overall on WorldScore.
- Storage is tunable: dropping plane resolution from 512 to 256 cuts memory from 73.9MB to 19.8MB with negligible quality loss. Limitations include reduced effective resolution as coverage grows and dependence on depth estimation quality.
Paper and code are public.
More from Multimodal
- MiniMax-H3 hailed as favorite open release of the year, runs fully local — AdinaYakup · 2026-10-23
- Michal Malewicz experiments with depth maps in his creative workflow — michalmalewicz · 2026-10-23
- One-Shot Tunable Style Transfer on Flux3 Via Reference Latent + Node — Winter_unmuted · 2026-10-23
- PhysVidBench accepted at EMNLP 2026: do video gen models grasp physical commonsense? — aykuterdemml · 2026-10-23
- EMNLP 2026 paper teaches multimodal LLMs to get cartoons via incongruity-resolution supervision — aykuterdemml · 2026-10-23
- A gorgeous Midjourney style: sref code 820221253 for dreamy lavender watercolor looks — umesh_ai · 2026-10-23