Improving Consistency Models with Generator-Augmented Flows
Abstract
Consistency models imitate the multi-step sampling of score-based diffusion in a single forward pass of a neural network. They can be learned in two ways: consistency distillation and consistency training. The former relies on the true velocity field of the corresponding differential equation, approximated by a pre-trained neural network. In contrast, the latter uses a single-sample Monte Carlo estimate of this velocity field. The related estimation error induces a discrepancy between consistency distillation and training that, we show, still holds in the continuous-time limit. To alleviate this issue, we propose a novel flow that transports noisy data towards their corresponding outputs derived from a consistency model. We prove that this flow reduces the previously identified discrepancy and the noise-data transport cost. Consequently, our method not only accelerates consistency training convergence but also enhances its overall performance. The code is available at: https://github.com/thibautissenhuth/consistency_GC.
Keywords
Cite
@article{arxiv.2406.09570,
title = {Improving Consistency Models with Generator-Augmented Flows},
author = {Thibaut Issenhuth and Sangchul Lee and Ludovic Dos Santos and Jean-Yves Franceschi and Chansoo Kim and Alain Rakotomamonjy},
journal= {arXiv preprint arXiv:2406.09570},
year = {2025}
}