jefequien/S2PD

Serial-to-Parallel Diffusion

Project Page · Code

DiT-B

Pixel-space models trained from scratch.

Dataset	Bidirectional	cDF	BC	S2PD
Conway	ckpt	ckpt	ckpt	ckpt
Chess	ckpt	ckpt	ckpt	ckpt
2048	ckpt	ckpt	ckpt	ckpt
Fifteen Puzzle	ckpt	ckpt	ckpt	ckpt
Tetris	ckpt	ckpt	ckpt	ckpt
Snake	ckpt	ckpt	ckpt	ckpt
Rubik’s Cube 3D	ckpt	ckpt	ckpt	ckpt
Double Pendulum	ckpt	ckpt	ckpt	ckpt
Three Body	ckpt	ckpt	ckpt	ckpt
Colliding Balls 2D	ckpt	ckpt	ckpt	ckpt
Colliding Balls 3D	ckpt	ckpt	ckpt	ckpt
Pong	ckpt	ckpt	ckpt	ckpt

Wan-5B-LoRA

Adapters requiring the Wan-5B backbone.

Dataset	Bidirectional	cDF	BC	S2PD
Kubric MOVi-A	ckpt	ckpt	ckpt	ckpt
Kubric MOVi-C	ckpt	ckpt	ckpt	ckpt
MPMWorlds	ckpt	ckpt	ckpt	ckpt
Rubik’s Cube Real	ckpt	ckpt	ckpt	ckpt

Block size ablations (DiT-B)

Pin-8 and Pin-64 train with their named block size and full-video blocks, weighted 80/20.

Dataset	Pin-8	Pin-64
Conway	ckpt	ckpt
Chess	ckpt	ckpt
2048	ckpt	ckpt
Fifteen Puzzle	ckpt	ckpt
Tetris	ckpt	ckpt
Snake	ckpt	ckpt
Rubik’s Cube 3D	ckpt	ckpt
Double Pendulum	ckpt	ckpt
Three Body	ckpt	ckpt
Colliding Balls 2D	ckpt	ckpt
Colliding Balls 3D	ckpt	ckpt
Pong	ckpt	ckpt

Token order ablations (DiT-B, Pin-8)

Dataset	Raster	Zigzag	Hilbert
Conway	ckpt	ckpt	ckpt
Chess	ckpt	ckpt	ckpt
2048	ckpt	ckpt	ckpt
Fifteen Puzzle	ckpt	ckpt	ckpt
Tetris	ckpt	ckpt	ckpt
Snake	ckpt	ckpt	ckpt
Rubik’s Cube 3D	ckpt	ckpt	ckpt
Double Pendulum	ckpt	ckpt	ckpt
Three Body	ckpt	ckpt	ckpt
Colliding Balls 2D	ckpt	ckpt	ckpt
Colliding Balls 3D	ckpt	ckpt	ckpt
Pong	ckpt	ckpt	ckpt

Download and usage instructions are in the code repository.

Datasets

• Kubric MOVi-A: 9,703 training and 250 validation clips.

• Kubric MOVi-C: 9,737 training and 250 validation clips.

• Rubik’s Cube Real: 5,000 training and 512 validation clips.

• Chess: 1 million training and 10,000 validation games.

Licenses

Original weights, adapters, Rubik’s Cube Real footage, and training curves: MIT. Lichess games: CC0. The mirrored Wan backbone, VAE, UMT5 encoder, and tokenizer: Apache 2.0.

S2PD Kubric renders: CC BY-SA 4.0; see the dataset cards for upstream attribution.

Citation

If you use S2PD in your research, please cite:

@misc{hu2026s2pd,
  title         = {{S2PD}: Serial-to-Parallel Diffusion for Physically and Logically Consistent Video Generation},
  author        = {Hu, Jeffrey and Olmeda Reino, Daniel and Tewari, Ayush},
  year          = {2026},
  eprint        = {2610.06847},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2610.06847}
}