PPMem: Plane Primitives as Persistent 3D Memory Let Video World Models Generate Long Coherent Videos
Robbyant-Research · hf · 2026-10-19
Tsinghua University, Ant Group and HKUST present PPMem, a video world model using compact plane primitives as persistent 3D memory for long coherent video generation.
- A small set of oriented planes (7 parameters each) captures dominant indoor/outdoor structures like walls, floors and facades
- A planar prediction head plus a Depth Token Fusion module lets each autoregressive chunk jointly produce RGB video and planar geometry, merged into a persistent plane memory that conditions later chunks via rendered depth
- Memory is 3 orders of magnitude more compact than per-pixel point clouds (100K primitives, 5 MB per scene), geometrically explicit with metric depth, normals and exportable surfaces
More from Multimodal
- Dreamina Canvas motion capture transfers character motion from reference video for $1.50/mo — SarahAnnabels · 2026-10-19
- Swap any video character with one reference image: full Recast workflow guide — aziz4ai · 2026-10-19
- AI Video Trend's Most Unique Take: Steve Carell Juggling a Soccer Ball in The Office — i____eat___ass · 2026-10-19
- Reddit debates whether MOSS 3 will repeat the Flux 3/SD3 censorship pattern — Chrono_Tri · 2026-10-19
- Two custom character LoRAs in one image? ComfyUI user seeks fix for feature bleed — badgaab · 2026-10-19
- Google's PaperBanana: Five Agents Auto-Generate Publication-Ready Paper Figures — aigclink · 2026-10-19