Drift Flow Matching

TL;DR

Drift Flow Matching (DFM) introduces a flexible generative framework that combines the efficiency of Drift Models with the iterative refinement capabilities of flow-based methods, adaptable to various quality and efficiency needs.

Contribution

DFM bridges the gap between one-step Drift Models and multi-step Flow Matching, enabling adaptive sampling with improved flexibility and efficiency.

Findings

DFM achieves effective generation across multiple tasks and datasets.

DFM allows refinement through multiple inference steps without sacrificing initial efficiency.

Experiments demonstrate DFM's generality and improved performance.

Abstract

Iterative generative models such as Flow Matching and Diffusion models have demonstrated strong test-time scaling behavior, where additional inference computation can improve generation quality. In contrast, Drift Models offer efficient one-step generation, but their direct generation paradigm limits such flexibility. In this work, we propose Drift Flow Matching (DFM), a framework that connects drifting generative modeling with flow-based iterative generation. DFM preserves the efficiency of direct transport maps while enabling generation to be refined through multiple inference steps when desired. This bridges the gap between one-step Drift Models and multi-step Flow Matching methods, and provides a novel generative paradigm that can adapt sampling computation to different quality--efficiency requirements. Extensive experiments across different tasks and datasets demonstrate the effectiveness and generality of the proposed framework.