Serial-to-Parallel Diffusion

Project Page · Code

DiT-B

Pixel-space models trained from scratch.

Wan-5B-LoRA

Adapters requiring the Wan-5B backbone.

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

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}
}