Tsinghua University releases PPMem, a plane-based memory for AI video models
Tsinghua University announced Persistent Planar Memory for Video World Models (PPMem), a new approach that stores 3D memory as compact plane primitives instead of dense point clouds. The system builds on the Wan 2.2‑5B video diffusion model and was trained on more than 80,000 clips from DL3DV and RealEstate10K, using depth data from MoGe‑2 and VGGT reconstructions. In tests, PPMem outperformed…
Key points
- PPMem uses plane primitives to store 3D memory, reducing 100k planes to 5MB, outperforming Gen3C, Voyager, VMEM, and SPMem on DL3DV and RealEstate10K.
- The system builds on Wan 2.2‑5B, training on 80,000 clips from DL3DV and RealEstate10K, with depth from MoGe‑2 and VGGT reconstructions.
- PPMem preserves scene consistency over 300–400 frames, achieving higher PSNR, LPIPS, CLIP and DINOv2 similarity than competitors, with a 0.27 margin at frame 300.
The paper argues that plane primitives provide a lightweight, structured memory that is orders of magnitude smaller than per‑pixel representations, enabling longer, more consistent AI‑generated video while reducing storage and inference costs. The authors suggest that this method could help future AI models retrieve and reference fixed assets more efficiently, though they acknowledge that the approach still relies on a “breadcrumb” strategy rather than a full RAG system.
Giving AI Video Some Remembrance of Things Past
Unite.AI · 20 October 2026
Loading the full article…
This text was published by Unite.AI and written by Martin Anderson. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Generative AI & Models
All →- OpenAI improves code migration for COBOL with GPT-6.1 Sol achieving 97.9% pass rate · 4 src
- Co-Existence released, uses Anthropic Opus 5.5 to chart 30 months of AI progress · 1 src
- MiniMax launches M3.1 Flash model at $0.15 per million input tokens with weights due October 23 · 5 src
- Could this be a Deep Blue moment for the cryptic crossword? · 1 src
- Ornith-ai/Ornith-1.5-122B (Qwen finetune) · 2 src
Comments
via GitHub Discussions