Search papers, labs, and topics across Lattice.
This paper investigates the dual parameterization of rectified-flow-based generative models, focusing on the distinct advantages of predicting instantaneous velocity versus data endpoints for denoising tasks. By analyzing how learning errors from these parameterizations influence generation performance, the authors introduce Self-Consistent Flow (SC-Flow), which integrates both approaches through a lightweight consistency loss. Experimental results show that SC-Flow enhances optimization stability and improves image generation quality compared to traditional rectified-flow models.
Unifying velocity and endpoint predictions reveals a method that significantly boosts generative model performance without major architectural changes.
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 behaviors, the mechanisms underlying their respective advantages remain to be underexplored, and how to combine them effectively is still unclear. In this work, we analyze how learning errors from different parameterizations affect the generation performance. We show that predicting the data endpoint has a clear training signal that stabilizes training, whereas predicting the velocity maintains stable sampling dynamics near the data manifold. Motivated by these insights, we propose Self-Consistent Flow (SC-Flow), a new method that unifies the benefits of both parameterizations. By employing a lightweight consistency loss, SC-Flow jointly trains a single network to predict both the local velocity and the data endpoint, and the consistency between the two predictions improves the model's performance. The method requires no major architectural changes and adds minimal computational overhead. Extensive experiments on image generation tasks demonstrate that SC-Flow substantially stabilizes optimization and improves the straightness of generation paths, leading to significant gains in generation quality over standard rectified-flow baselines.
