English

TRACT: Denoising Diffusion Models with Transitive Closure Time-Distillation

Machine Learning 2023-03-09 v1 Computer Vision and Pattern Recognition

Abstract

Denoising Diffusion models have demonstrated their proficiency for generative sampling. However, generating good samples often requires many iterations. Consequently, techniques such as binary time-distillation (BTD) have been proposed to reduce the number of network calls for a fixed architecture. In this paper, we introduce TRAnsitive Closure Time-distillation (TRACT), a new method that extends BTD. For single step diffusion,TRACT improves FID by up to 2.4x on the same architecture, and achieves new single-step Denoising Diffusion Implicit Models (DDIM) state-of-the-art FID (7.4 for ImageNet64, 3.8 for CIFAR10). Finally we tease apart the method through extended ablations. The PyTorch implementation will be released soon.

Keywords

Cite

@article{arxiv.2303.04248,
  title  = {TRACT: Denoising Diffusion Models with Transitive Closure Time-Distillation},
  author = {David Berthelot and Arnaud Autef and Jierui Lin and Dian Ang Yap and Shuangfei Zhai and Siyuan Hu and Daniel Zheng and Walter Talbott and Eric Gu},
  journal= {arXiv preprint arXiv:2303.04248},
  year   = {2023}
}
R2 v1 2026-06-28T09:06:31.129Z