2 diffusion transformer.
Plane primitives serve as compact persistent memory for video world models
PPMem uses a few hundred thousand plane primitives to maintain geometric coherence across hundreds of generated frames under large viewpoint changes.
Chinese Tech
Yuze He · Bin Tan · Zelin Gao · Yujun Shen · Yong-Jin Liu · Nan Xue · +1 more
Tsinghua University · Ant Group · HKUST
Research Digest··2 min read
The authors present PPMem, a video world model that augments a diffusion transformer with a planar prediction head and a Depth Token Fusion module.
§