Honeycomb: Constant-Size Scene Memory Representation for Video World Models
Honeycomb keeps video world-model scene memory constant using six HexMemory planes.
Honeycomb is a video world model that stores scene features in HexMemory, a fixed-size representation of six spatial and spatiotemporal planes. A feed-forward writer maps each new chunk into plane features, warps prior planes without changing their dimensions, and fuses them with confidence-weighted pooling plus a learned residual. A reader conditions later generation on retrieved latents, processing only the new chunk rather than the full history. Experiments on WorldScore and RealEstate10K report strong generation quality and revisit consistency with constant memory.
- HexMemory stores scene features in six fixed-size spatial and spatiotemporal planes.
- A feed-forward writer fuses new chunks via warping, pooling, and residual correction.
- WorldScore and RealEstate10K show strong quality and robust revisit consistency.
- Feature storage stays constant and the writer skips full-history reprocessing.
- AI models
- Honeycomb
From the original article
1 min read at huggingface.co
Video world models require persistent scene memory to maintain consistency during long-horizon video generation. Existing spatial memories accumulate RGB observations or latent features, increasing storage requirements as generation proceeds.
By Jack Wei Lun Shi, Kaichen Zhou, Haoyu Chen · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.