AlphaFlow: Understanding and Improving MeanFlow Models

TL;DR

AlphaFlow analyzes why MeanFlow works by showing its objective splits into trajectory flow matching and trajectory consistency, whose gradients conflict during joint optimization. It proposes α-Flow, a curriculum-based loss that anneals from flow matching to MeanFlow, unifying one-, few-, and many-step flow models and improving convergence for from-scratch image generation. The approach yields strong empirical gains on ImageNet-1K 256^2 with vanilla DiT backbones, achieving state-of-the-art 1-NFE and 2-NFE FID scores (2.58 and 2.15) and providing extensive ablations and reproducibility details. These results highlight the practical impact of curriculum learning in complex diffusion-flow objectives and offer a scalable path to faster, higher-fidelity few-shot 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 $α$-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, $α$-Flow disentangles the conflicting objectives, and achieves better convergence. When trained from scratch on class-conditional ImageNet-1K 256x256 with vanilla DiT backbones, $α$-Flow consistently outperforms MeanFlow across scales and settings. Our largest $α$-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).

Figures (25)

• Figure 1: Uncurated samples (seeds $1$-$8$) from the DiT-XL/2 model for MeanFlow MeanFlow and $\alpha$-Flow (our proposed method) produced with 1 (upper) and 2 (lower) sampling steps for ImageNet-1K $256^2$.
• Figure 2: MeanFlow training analysis. (a) Shows the cosine similarity between the gradients of two loss pairs ($\nabla \mathcal{L}_{\mathtt{TC_c}}\xspace$ vs. $\nabla \mathcal{L}_{\mathtt{TFM}}\xspace$ and $\nabla \mathcal{L}_{\mathtt{TC_c}}\xspace$ vs. $\nabla \mathcal{L}_{\mathtt{FM'}}\xspace$) throughout training. (b) Evaluated $\mathcal{L}_{\mathtt{TFM}}\xspace$ when MeanFlow trained with 0% and 75% of $\mathcal{L}_{\mathtt{FM'}}\xspace$. (c) Evaluated $\mathcal{L}_{\mathtt{TC_c}}\xspace$ when MeanFlow trained with 0% and 75% of $\mathcal{L}_{\mathtt{FM'}}\xspace$.
• Figure 3: Comparison of training trajectories for various few-step diffusion and flow-based models.
• Figure 4: Comparing ODE vs consistency sampling for MeanFlow and $\alpha$-Flow models.
• Figure 5: Visualization of consistency step ratio schedule.

Theorems & Definitions (5)

• Definition 1
• Theorem 1
• proof
• proof : Proof of \ref{['thm:unified-loss']}
• proof : Proof of equivalence with consistency model