Improving Consistency Models with Generator-Augmented Flows

Thibaut Issenhuth, Sangchul Lee, Ludovic Dos Santos, Jean-Yves Franceschi, Chansoo Kim, Alain Rakotomamonjy
Proceedings of the 42nd International Conference on Machine Learning, PMLR 267:26586-26610, 2025.

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.

Cite this Paper

BibTeX
@InProceedings{pmlr-v267-issenhuth25a, title = {Improving Consistency Models with Generator-Augmented Flows}, author = {Issenhuth, Thibaut and Lee, Sangchul and Dos Santos, Ludovic and Franceschi, Jean-Yves and Kim, Chansoo and Rakotomamonjy, Alain}, booktitle = {Proceedings of the 42nd International Conference on Machine Learning}, pages = {26586--26610}, year = {2025}, editor = {Singh, Aarti and Fazel, Maryam and Hsu, Daniel and Lacoste-Julien, Simon and Berkenkamp, Felix and Maharaj, Tegan and Wagstaff, Kiri and Zhu, Jerry}, volume = {267}, series = {Proceedings of Machine Learning Research}, month = {13--19 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v267/main/assets/issenhuth25a/issenhuth25a.pdf}, url = {https://proceedings.mlr.press/v267/issenhuth25a.html}, 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.} }
Endnote
%0 Conference Paper %T Improving Consistency Models with Generator-Augmented Flows %A Thibaut Issenhuth %A Sangchul Lee %A Ludovic Dos Santos %A Jean-Yves Franceschi %A Chansoo Kim %A Alain Rakotomamonjy %B Proceedings of the 42nd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2025 %E Aarti Singh %E Maryam Fazel %E Daniel Hsu %E Simon Lacoste-Julien %E Felix Berkenkamp %E Tegan Maharaj %E Kiri Wagstaff %E Jerry Zhu %F pmlr-v267-issenhuth25a %I PMLR %P 26586--26610 %U https://proceedings.mlr.press/v267/issenhuth25a.html %V 267 %X 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.
APA
Issenhuth, T., Lee, S., Dos Santos, L., Franceschi, J., Kim, C. & Rakotomamonjy, A.. (2025). Improving Consistency Models with Generator-Augmented Flows. Proceedings of the 42nd International Conference on Machine Learning, in Proceedings of Machine Learning Research 267:26586-26610 Available from https://proceedings.mlr.press/v267/issenhuth25a.html.

Related Material