FAE — Scalable Normalizing Flows

TL;DR

FAE adapts powerful pretrained visual encoders into generation-friendly low-dimensional latents using as little as a single attention layer — reaching state-of-the-art ImageNet-256 FID for both diffusion and normalizing-flow generators.

Abstract

Visual generative models (e.g., diffusion models) typically operate in compressed latent spaces to balance training efficiency and sample quality. In parallel, there has been growing interest in leveraging high-quality pre-trained visual representations, either by aligning them inside VAEs or directly within the generative model. However, adapting such representations remains challenging due to fundamental mismatches between understanding-oriented features and generation-friendly latent spaces. Representation encoders benefit from high-dimensional latents that capture diverse hypotheses for masked regions, whereas generative models favor low-dimensional latents that must faithfully preserve injected noise. This discrepancy has led prior work to rely on complex objectives and architectures. In this work, we propose FAE (Feature Auto-Encoder), a simple yet effective framework that adapts pre-trained visual representations into low-dimensional latents suitable for generation using as little as a single attention layer, while retaining sufficient information for both reconstruction and understanding. The key is to couple two separate deep decoders: one trained to reconstruct the original feature space, and a second that takes the reconstructed features as input for image generation. FAE is generic; it can be instantiated with a variety of self-supervised encoders (e.g., DINO, SigLIP) and plugged into two distinct generative families: diffusion models and normalizing flows. Across class-conditional and text-to-image benchmarks, FAE achieves strong performance. For example, on ImageNet 256x256, our diffusion model with CFG attains a near state-of-the-art FID of 1.29 (800 epochs) and 1.70 (80 epochs). Without CFG, FAE reaches the state-of-the-art FID of 1.48 (800 epochs) and 2.08 (80 epochs), demonstrating both high quality and fast learning.

Method

Overview figure from the paper — see the linked paper for full details.

Key Contributions

Coupling two deep decoders: one reconstructs the original feature space, while a second turns the reconstructed features into images.

A one-attention-layer adapter that turns understanding-oriented features into low-dimensional, generation-ready latents.

Generic across encoders (DINO, SigLIP) and generative families, reaching SOTA ImageNet-256 FID (e.g. 1.48 without CFG).

BibTeX

@inproceedings{gao2025one,
  title     = {One Layer Is Enough: Adapting Pretrained Visual Encoders for Image Generation},
  author    = {Gao, Yuan and Chen, Chen and Chen, Tianrong and Gu, Jiatao},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Findings},
  year      = {2026}
}