PPMem stores video world-model scenes as compact planar memory
Original title: Persistent Planar Memory for Video World Models
About 100K plane primitives, roughly 5MB per scene, stay coherent across hundreds of frames under large viewpoint changes.
PPMem is a video world model that keeps persistent 3D memory as oriented plane primitives rather than per-pixel point clouds. A video diffusion transformer with a planar prediction head and Depth Token Fusion jointly emits RGB video and planar geometry, which merge into memory that conditions later chunks via rendered depth. On DL3DV and RealEstate10K, the paper reports better video quality than prior 3D-grounded world models while also producing metric depth, normals, and exportable surfaces.
- Each plane primitive uses 7 parameters.
- Memory is described as three orders of magnitude more compact than per-pixel point clouds.
- Geometry is jointly predicted by the shared backbone and refined as generation proceeds.
From the original article
1 min read at huggingface.co
Video world models that generate and explore environments at scale need a persistent 3D memory of what they have produced. Plane primitives offer a natural candidate: a compact set of oriented planes captures the dominant structural elements (e. g.
By Yuze He, Bin Tan, Zelin Gao · 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.