End-to-End Single-Step Flow Matching via Direct Models

ICLR 2026 Conference Submission16710 Authors

Keywords: Flow matching, single step generative models

Abstract: We introduce Direct Models , a flow-matching framework that enables single-step generation by learning a direct mapping from initial noise $x_0$ to all intermediate latent states along the generative trajectory. Our method is trained end-to-end and does not rely on multi-stage distillation. It leverages a progressive learning scheme where the mapping from $x_0$ to $x_{t + \delta t}$ is composed as an update from $x_0$ to $x_t$ plus the velocity at time $t$. This formulation allows the model to learn the entire trajectory in a recursive, data-consistent manner while maintaining computational efficiency. Experimentally, we show that Direct Models achieves state-of-the-art sample quality among single-step flow-matching methods.

Primary Area: generative models

Submission Number: 16710