AlphaFlow: Understanding and Improving MeanFlow Models
Huijie Zhang, Aliaksandr Siarohin, Willi Menapace, Michael Vasilkovsky, Sergey Tulyakov, Qing Qu, Ivan Skorokhodov
2025-10-24
Summary
This paper investigates a new method for creating images called MeanFlow, which is good at generating images with only a few steps, but researchers weren't sure *why* it worked so well. The paper aims to understand what makes MeanFlow successful and improve it.
What's the problem?
The researchers found that MeanFlow's process can be broken down into two parts that actually work against each other. Imagine trying to steer a car while simultaneously trying to make it go straight – those two goals conflict! This conflict slows down the learning process and makes it harder for the model to create high-quality images. Essentially, the different parts of the method were pulling in opposite directions, causing optimization issues.
What's the solution?
To fix this, the researchers created a new approach called alpha-Flow. Alpha-Flow is a more general method that includes MeanFlow as one possibility, but also allows for a smoother learning process. They trained the model gradually, starting with a simpler goal (trajectory flow matching) and slowly transitioning to the more complex goal of MeanFlow. This 'curriculum' approach helps the model learn each part effectively before combining them, avoiding the conflict and leading to faster and better results.
Why it matters?
This work is important because it provides a better understanding of how MeanFlow works and, more importantly, it improves upon it. The alpha-Flow method achieves better image quality, as measured by a metric called FID score, than previous methods when creating images from scratch. This means we can now generate even more realistic and detailed images with fewer steps, which is a significant advancement in the field of image generation.
Abstract
MeanFlow has recently emerged as a powerful framework for few-step generative modeling trained from scratch, but its success is not yet fully understood. In this work, we show that the MeanFlow objective naturally decomposes into two parts: trajectory flow matching and trajectory consistency. Through gradient analysis, we find that these terms are strongly negatively correlated, causing optimization conflict and slow convergence. Motivated by these insights, we introduce alpha-Flow, a broad family of objectives that unifies trajectory flow matching, Shortcut Model, and MeanFlow under one formulation. By adopting a curriculum strategy that smoothly anneals from trajectory flow matching to MeanFlow, alpha-Flow disentangles the conflicting objectives, and achieves better convergence. When trained from scratch on class-conditional ImageNet-1K 256x256 with vanilla DiT backbones, alpha-Flow consistently outperforms MeanFlow across scales and settings. Our largest alpha-Flow-XL/2+ model achieves new state-of-the-art results using vanilla DiT backbones, with FID scores of 2.58 (1-NFE) and 2.15 (2-NFE).