Paper ID: 2406.09570
Improving Consistency Models with Generator-Induced Flows
Thibaut Issenhuth, Sangchul Lee, Ludovic Dos Santos, Jean-Yves Franceschi, Chansoo Kim, Alain Rakotomamonjy
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 the currently trained model --~as a proxy of the true flow. Our empirical findings demonstrate that this approach mitigates the previously identified discrepancy. Furthermore, we present theoretical and empirical evidence indicating that our generator-induced flow surpasses dedicated optimal transport-based consistency models in effectively reducing 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: this https URL
Submitted: Jun 13, 2024