Issenhuth, Thibaut Lee, Sangchul Dos Santos, Ludovic Franceschi, Jean-Yves Kim, Chan soo Rakotomamonjy, Alan 2025-11-28T04:30:42Z 2025-11-28T04:30:42Z 2025-11-28 2025-07 https://pubs.kist.re.kr/handle/201004/153710 https://proceedings.mlr.press/v267/issenhuth25a.html 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. JMLR-JOURNAL MACHINE LEARNING RESEARCH Improving Consistency Models with Generator-Augmented Flows Conference 1 42nd International Conference on Machine Learning-ICML-Annual 2025, pp.26586 - 26610 42nd International Conference on Machine Learning-ICML-Annual 2025 26586 26610 CN Vancouver, CANADA 2025-07-13 INTERNATIONAL CONFERENCE ON MACHINE LEARNING