Jiajun Fan

Fine-tuning Flow Matching Generative Models with Intermediate Feedback

In brief

What the paper does.

We present AC-Flow, a robust actor-critic framework for fine-tuning flow matching generative models with intermediate feedback. Key innovations include reward shaping for stable intermediate value learning, a dual-stability mechanism combining advantage clipping with critic warm-up, and a scalable generalized critic weighting scheme with Wasserstein regularization. Achieves SOTA text-to-image alignment on Stable Diffusion 3 without compromising diversity or stability.

Published in
arXiv preprint
Paper
arxiv.org/abs/2510.18072
Project page
jiajunfan.com/projects/ac-flow/

Cite

Cite the paper.

@misc{fan2025finetuningflowmatchinggenerative,
  title={Fine-tuning Flow Matching Generative Models with Intermediate Feedback},
  author={Jiajun Fan and Chaoran Cheng and Shuaike Shen and Xiangxin Zhou and Ge Liu},
  year={2025},
  eprint={2510.18072},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2510.18072}
}