TerryYu/DOGTrim

Rethinking Test Time Scaling for Flow-Matching Generative Models

About

This repository contains the models and configuration for our paper Rethinking Test Time Scaling for Flow-Matching Generative Models.

After analyzing the limitations of existing methods on ODE flow-matching models, we propose:

DOG-Trim: Diversity enhanced Order aligned Global flow Trimming

Qualitative examples using Flux1.dev:

Citation

If you find this work useful, please consider citing our arXiv preprint.

@article{yu2026RethinkTTS,
  title={Rethinking Test Time Scaling for Flow-Matching Generative Models},
  author={Yu, Qingtao and Song, Changlin and Sun, Minghao and Yu, Zhengyang and Verma, Vinay Kumar and Roy, Soumya and Negi, Sumit and Li, Hongdong and Campbell, Dylan},
  journal={arXiv preprint arXiv:2511.22242},
  year={2026}
}