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

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, 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 variational score distillation (VSD) while ensuring global trajectory consistency. We further prove the Mean Flow Matching Theorem, establishing that matching expected average velocities is sufficient for strict distribution alignment. Empirically, on challenging high-dimensional manifolds including 4D occupancy forecasting and text-to-image generation, MFD achieves state-of-the-art performance, enabling high-fidelity single-step generation.

Modern generative AI models can produce high-quality images and other complex outputs, but many of them require many repeated computation steps before producing a final result. This makes them slow and expensive to use in real-time applications. This paper introduces Mean Flow Distillation, a method for making flow matching generative models much faster. Instead of asking a smaller student model to imitate every tiny step of a large teacher model, our method teaches it to match the teacher’s average movement over time. This provides a smoother and more reliable learning signal, reducing training instability. As a result, the distilled model can generate high-quality outputs in a single step while preserving much of the quality of the original multi-step model. We validate the method on challenging tasks including text-to-image generation and 4D occupancy forecasting.