README.md

February 4, 2026 · View on GitHub

RCGM: Any-step Generation via N-th Order Recursive Consistent Velocity Field Estimation

Peng Sun^1,2 · Tao Lin¹

¹Westlake University ²Zhejiang University

:robot: Models :page_facing_up: Paper :label: BibTeX

Official PyTorch implementation of RCGM: :trophy: A unified, scalable, and stable generative framework.

Generated samples from our RCGM-tuned Qwen-Image-20B model (NFE=8), which achieved a GenEval score of 0.87 and required only about 80 H800 GPU hours for tuning.

:construction: TODOs

[✅] Release the paper.

[ ] Release trained models.

[ ] Update the paper.

[ ] Release training and inference code.

:label: Bibliography

If you find this repository helpful for your project, please consider citing our work:

@inproceedings{sun2026anystep,
  title     = {Any-step Generation via N-th Order Recursive Consistent Velocity Field Estimation},
  author    = {Sun, Peng and Lin, Tao},
  booktitle = {The Fourteenth International Conference on Learning Representations},
  year      = {2026},
  url       = {https://openreview.net/forum?id=GnawtLKGkP}
}

:page_facing_up: License

Apache License 2.0 - See LICENSE for details.