A Consistent Flow Model Learning Both Where to Go and How to Move

Xu Han, Jiajing Hu, Liping Liu

Keywords: GenAI, Generative model, Diffusion, Flow, Ode

TL;DR: BiFlow is a plug-and-play single-head rectified flow that jointly learns velocity and destination with a consistency tie, keeps computation the same, and improves FID via a simple time-scheduled drift switch.

Abstract: Under the framework of rectified flow, generative neural networks are trained to a single vector field specified by straight paths between data samples and random samples from the prior. This work reveals equivalent two forms of the learning problem, with each form setting up a fitting target for the neural network. We then introduce a new model, BiFlow, which is trained with two complementary targets-the velocity of the vector field and the likely destination of ODE paths-thereby endows the model with local sensitivity (how to move now) and global awareness (where the path should end). The new design uses a single head with a binary mode flag to output either prediction, plus a lightweight consistency loss that ties them together. It drops into existing pipelines: no architectural overhaul, the usual conditioning, therefore adding no extra cost to the generation procedure. Our experiment shows that the new design stabilizes optimization and improves the straightness of generation paths. On two image generation tasks, BiFlow substantially improves generation quality over rectified flow.

Primary Area: generative models

Submission Number: 4170