FREPix: Frequency-Heterogeneous Flow Matching for Pixel-Space Image Generation

TL;DR

FREPix introduces a frequency-heterogeneous flow matching framework that explicitly decomposes image generation into low- and high-frequency components, improving pixel-space image synthesis.

Contribution

It proposes a novel frequency-aware approach that explicitly models and trains low- and high-frequency components separately for pixel-space image generation.

Findings

Achieves 1.91 FID at 256x256 resolution on ImageNet.

Outperforms existing pixel-space models in low-NFE regimes.

Effectively separates frequency components for coarse-to-fine generation.

Abstract

Pixel-space diffusion has re-emerged as a promising alternative to latent-space generation because it avoids the representation bottleneck introduced by VAEs. Yet most existing methods still treat image generation as a frequency-homogeneous process, overlooking the distinct roles and learning dynamics of low- and high-frequency components. To address this, we propose FREPix, a FREquency-heterogeneous flow matching framework for Pixel-space image generation. FREPix explicitly decomposes generation into low- and high-frequency components, assigns them separate transport paths, predicts them with a factorized network, and trains them with a frequency-aware objective. In this way, coarse-to-fine generation becomes an explicit design principle rather than an implicit behavior. On ImageNet class-to-image generation, FREPix achieves competitive results among pixel-space generation models, reaching 1.91 FID at $256\times256$ and 2.38 FID at $512\times512$, with particularly strong behavior in the low-NFE regime.