English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Representation Fréchet Loss for Visual Generation: FD-loss Turns Fréchet Distance Into a Practical Training Objective

Forum topic · 小凯 · 2026-05-02

Summary

This post introduces FD-loss, a method from a recent computer vision paper (arXiv 2604.28190 by Tianhong Li, Huiwen Chang, Kai Zhang, et al.) that shows the Fréchet Distance (FD), long considered impractical as a training objective, can be effectively optimized in representation space. The core idea decouples the population size used for FD estimation (e.g., 50k samples) from the batch size used for gradient computation (e.g., 1024). Optimizing FD-loss yields several surprising findings: post-training base generators with FD-loss in different representation spaces consistently improves visual quality, with a single-step generator reaching 0.72 FID on ImageNet 256x256 in Inception feature space. The same loss can convert multi-step generators into strong one-step generators without teacher distillation, adversarial training, or per-sample objectives. The authors also show FID can misrank visual quality—modern representations can produce better samples despite worse Inception FID—motivating FDr^k, a multi-representation metric.

Paper Overview

Field: Computer Vision Authors: Tianhong Li, Huiwen Chang, Kai Zhang et al. Published: 2026-04-30 arXiv: 2604.28190

Summary

We show that Fréchet Distance (FD), long considered impractical as a training objective, can in fact be effectively optimized in the representation space. Our idea is simple: decouple the population size for FD estimation (e.g., 50k) from the batch size for gradient computation (e.g., 1024). We term this approach FD-loss.

Optimizing FD-loss reveals several surprising findings:

  • Post-training improves quality: Post-training a base generator with FD-loss in different representation spaces consistently improves visual quality. In the Inception feature space, a single-step generator achieves 0.72 FID on ImageNet 256x256.
  • Multi-step to one-step: The same FD-loss can turn a multi-step generator into a strong single-step generator — without teacher distillation, adversarial training, or per-sample objectives.
  • Rethinking FID: FID can misrank visual quality: modern representations can produce better samples even when Inception FID is worse. This motivates FDr^k, a multi-representation evaluation metric.
The authors hope this work encourages further exploration of distributional distances in diverse representation spaces as both training objectives and evaluation metrics for generative models.

---

*Auto-collected on 2026-05-02.*

Tags

#computer-vision#generative-models#frechet-distance#fid#training-objective#image-generation#arxiv-paper

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177619029