AdvFD: Boosting Visual Generation via Adversarial Fréchet Distance Loss

Forum topic · 小凯 · 2026-08-13

Summary

AdvFD (Adversarial Fréchet Distance) is a proposed post-training objective for visual generative models, addressing the problem of Fréchet hacking, where optimizing a Fréchet Distance (FD) loss improves the target metric but visual quality and Fréchet alignment in other feature spaces stagnate or degrade. The authors attribute this to the static pretrained feature spaces used by existing FD losses, which offer incomplete, fixed views of the gap between real and generated distributions. AdvFD augments the static FD objective with a learnable, adversarially calibrated representation: an adversary maximizes the Fréchet discrepancy between real and generated samples, while the generator minimizes it in this adaptive feature space. To prevent the adversary from trivially inflating the objective via feature amplification, the method introduces real-feature whitening, which normalizes scale and covariance geometry and stabilizes the min-max optimization. Experiments show consistent improvements in one-step generator post-training across JiT and pMF backbones and various model scales. Paper: arXiv:2608.11205 by Mingju Gao, Jingkai Zhou, Kun Gai, Changqian Yu, and Hao Tang.

AdvFD: Boosting Visual Generation via Adversarial Fréchet Distance Loss

Field: Computer Vision Authors: Mingju Gao, Jingkai Zhou, Kun Gai, Changqian Yu, Hao Tang arXiv: 2608.11205

Abstract (translated)

Fréchet distance has recently emerged as an effective distribution-level objective for generator post-training, complementing the conventional sample-level diffusion and flow-matching losses. However, directly optimizing Fréchet objectives can cause Fréchet hacking: the target metrics keep improving, but visual quality and Fréchet alignment in other feature spaces may stagnate or deteriorate.

The authors attribute this failure to the static pretrained feature spaces used by existing Fréchet losses, which provide incomplete and fixed views of the differences between real and generated distributions.

Method

To address this limitation, they propose Adversarial Fréchet Distance (AdvFD), which complements the static representation targets in FD-Loss with a calibrated adversarially learned representation:

A learnable representation adversarially *maximizes* the Fréchet discrepancy between real and generated samples.
The generator *minimizes* the same discrepancy in this resulting adaptive feature space.
To prevent the adversarial representation from trivially inflating the objective through feature amplification, they introduce real-feature whitening, which normalizes the scale and covariance geometry of real features and stabilizes the min-max optimization.

Results

Extensive experiments show that AdvFD consistently improves one-step generator post-training across JiT and pMF backbones and across different model scales.

---

*Auto-collected on 2026-08-13.*

AdvFD: Boosting Visual Generation via Adversarial Fréchet Distance Loss

Summary

AdvFD: Boosting Visual Generation via Adversarial Fréchet Distance Loss

Abstract (translated)

Method

Results

Tags