Mean Flow Distillation: A Robust and Stable Distillation Framework 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 generative models. Flow Matching achieves strong performance across generative tasks, but its ODE-based iterative sampling introduces heavy computational overhead that 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 generation quality. The authors theoretically show that MFD acts as a temporal low-pass filter that suppresses the high-frequency optimization noise inherent in Variational Score Distillation (VSD) while ensuring global trajectory consistency. They further prove a mean flow matching theorem: matching the expected average velocity is sufficient to achieve 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 enables high-fidelity single-step generation. Paper: arXiv 2606.11155.

Overview

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

Abstract (translated)

Mean Flow Distillation (MFD) is a distillation framework purpose-built for Flow Matching models. The authors theoretically prove 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. They also establish a mean flow matching theorem: matching the expected average velocity is sufficient to achieve 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.

Original Abstract (excerpt)

> 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, we propose Mean Flow Distillation (MFD), a novel distillation framework tailored for flow matching models. We theoretically demonstrate that MFD acts as a temporal low-pass filter, effectively suppressing the high-frequency optimization noise inherent in variance score distillation...

Full paper: https://arxiv.org/abs/2606.11155

--- *Auto-collected on 2026-06-11*