Robbyant-Research/PPMem

Persistent Planar Memory for Video World Models

Yuze He^1,2, Bin Tan², Zelin Gao^3,2, Yujun Shen², Yong-Jin Liu^1,*, Nan Xue^2,*, Yinghao Xu^3,2

¹Tsinghua University    ²Ant Group    ³HKUST
^*Corresponding authors.
NeurIPS 2026

Project page · Paper · Code

PPMem generates videos from a single image and a camera trajectory. It accumulates reconstructed 3D planes in a persistent memory and projects them into new views to condition subsequent video segments.

Model

PPMem extends Wan 2.2 TI2V-5B with depth-token fusion, a feature decoder, a plane head, and rank-256 attention LoRA.

Property	Value
Base model	Wan 2.2 TI2V-5B
Weight format	Sharded safetensors
Precision	bfloat16

Checkpoint files

Path	Contents
model.safetensors.index.json	Tensor-to-shard mapping.
model-0000x-of-00003.safetensors	PPMem weights.
vae/	Wan 2.2 VAE.
text_encoder/	Wan 2.2 text encoder.
tokenizer/	Wan 2.2 tokenizer.
assets/	Demonstration video and its poster, used by this page.

Citation

@article{he2026ppmem,
  title   = {Persistent Planar Memory for Video World Models},
  author  = {He, Yuze and Tan, Bin and Gao, Zelin and Shen, Yujun
             and Liu, Yong-Jin and Xue, Nan and Xu, Yinghao},
  journal = {arXiv preprint arXiv:2610.23795},
  year    = {2026}
}

License

PPMem is licensed under Apache 2.0. The bundled VAE, text encoder, and tokenizer are from Wan 2.2 TI2V-5B, also licensed under Apache 2.0.