Li-Ping Liu

Self-Consistent Flow: Unifying Velocity and Endpoint Prediction for Rectified Flow Models

Xu Han, Jiajing Hu, Li-Ping Liu

In rectified-flow-based generative models, the neural network can be trained to predict two different targets, such as the instantaneous velocity or the data endpoint, to perform denoising. Although prior work shows that these parameterizations lead to different empirical behavio…Read more