Mean Flows for One-step Generative Modeling

2025-05-19

AI Summary

To address the performance gap between single-step and multi-step generative models, this paper proposes the Average Flow Velocity Modeling framework: it replaces the instantaneous velocity in conventional flow matching with the time-averaged flow velocity, constraining neural network training via a rigorously derived average–instantaneous velocity identity. This approach is the first to adopt average velocity as the core representation, enabling end-to-end single-step flow matching without pretraining, knowledge distillation, or curriculum learning, while supporting 1-NFE (number of function evaluations) forward sampling. Trained from scratch on ImageNet 256×256, the model achieves an FID of 3.43—setting the new state-of-the-art for single-step diffusion and flow-based models—and substantially narrows the performance gap with multi-step counterparts.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Deep Generative Models & Autoencoders
• Natural Language Processing: Generation

Application Category

• User Modeling, Personalization and Recommendation: Federated recommendation systems and personalization
• Social Networks and Social Media: Generative AI / large language models and their impact on social systems
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

We propose a principled and effective framework for one-step generative modeling. We introduce the notion of average velocity to characterize flow fields, in contrast to instantaneous velocity modeled by Flow Matching methods. A well-defined identity between average and instantaneous velocities is derived and used to guide neural network training. Our method, termed the MeanFlow model, is self-contained and requires no pre-training, distillation, or curriculum learning. MeanFlow demonstrates strong empirical performance: it achieves an FID of 3.43 with a single function evaluation (1-NFE) on ImageNet 256x256 trained from scratch, significantly outperforming previous state-of-the-art one-step diffusion/flow models. Our study substantially narrows the gap between one-step diffusion/flow models and their multi-step predecessors, and we hope it will motivate future research to revisit the foundations of these powerful models.

Problem

• Proposes a framework for one-step generative modeling
• Introduces average velocity to characterize flow fields
• Achieves state-of-the-art performance on ImageNet 256x256

Innovation

• Introduces average velocity for flow fields
• Derives identity between average and instantaneous velocities
• Self-contained MeanFlow model without pre-training