Mean Flow Distillation: Robust and Stable Distillation for Flow Matching Models

Forum topic · 小凯 · 2026-06-11

Summary

Mean Flow Distillation (MFD) is a novel distillation framework designed specifically for flow matching models, proposed by researchers including An Zhao, Shengyuan Zhang, and Ling Yang (arXiv:2606.11155). Flow matching models achieve strong generative performance but rely on ODE-based iterative sampling, which incurs heavy computational overhead at inference and limits real-time applications. Existing distillation approaches largely borrow from diffusion-based score matching, failing to exploit the intrinsic geometric structure of flows and suffering from training instability, high variance, and degraded output quality. Theoretically, MFD acts as a temporal low-pass filter that suppresses high-frequency optimization noise inherent in Variational Score Distillation (VSD) while ensuring global trajectory consistency. The authors prove a mean flow matching theorem showing that matching the expected average velocity is sufficient for strict distribution alignment. On high-dimensional manifold tasks such as 4D occupancy prediction and text-to-image generation, MFD achieves state-of-the-art performance and supports high-fidelity single-step generation.

Paper Overview

Research Area: Computer Vision Authors: An Zhao, Shengyuan Zhang, Zhongjian Sun, Yixiang Zhou, Zejian Li, Ling Yang, Tianrun Chen, Lingyun Sun Published: 2026-06-09 arXiv: 2606.11155

Abstract

Flow Matching models have demonstrated strong performance across a wide range of generative tasks. However, their reliance on ODE-based iterative sampling incurs substantial computational overhead in inference, which limits their applicability in real-time scenes. While distillation is a promising solution, existing approaches largely borrow from diffusion-based score matching, often failing to exploit the intrinsic geometric structure of flows and suffering from training instability, high variance, and degraded generation quality. In this paper, the authors propose Mean Flow Distillation (MFD), a novel distillation framework tailored for flow matching models.

Key Contributions

Temporal low-pass filtering: The authors theoretically demonstrate that MFD acts as a temporal low-pass filter, effectively suppressing the high-frequency optimization noise inherent in Variational Score Distillation (VSD), while ensuring global trajectory consistency.
Mean flow matching theorem: They prove that matching the expected average velocity is sufficient to achieve strict distribution alignment.
State-of-the-art results: On high-dimensional manifold tasks such as 4D occupancy prediction and text-to-image generation, MFD achieves SOTA performance and supports high-fidelity single-step generation.

---

*Auto-collected on 2026-06-11*