Paper Overview
- Field: Computer Vision (CV)
- Authors: Mingju Gao, Jingkai Zhou, Kun Gai, Changqian Yu, Hao Tang
- Released: 2026-08-11
- arXiv: 2608.11205
- Problem identified: Direct optimization of Fréchet distance losses leads to "Fréchet hacking," where metric scores rise but visual fidelity and cross-feature-space alignment do not.
- Root cause: Existing Fréchet losses rely on static pretrained feature spaces that give an incomplete, fixed view of real-vs-generated distributional gaps.
- Method (AdvFD): Adds an adversarially learned representation to the static Fréchet objective. The representation is trained to maximize real-vs-generated Fréchet divergence, while the generator is trained to minimize divergence in this adaptive space.
- Stabilization technique: Real-feature whitening normalizes scale and covariance geometry to prevent adversarial features from inflating the objective via feature amplification, stabilizing the min-max optimization.
- Empirical results: Consistent improvements in one-step post-training generation on JiT and pMF backbones across multiple model scales.
- arXiv link: https://arxiv.org/abs/2608.11205
- Auto-collected: 2026-08-13
Chinese Summary (Translated)
Fréchet distance has recently emerged as an effective distribution-level objective for generator post-training, complementing conventional sample-level diffusion and flow-matching losses. However, directly optimizing Fréchet objectives can cause *Fréchet hacking* — the target metrics keep improving, while 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 an incomplete and fixed view of the differences between real and generated distributions.
To address this limitation, the paper proposes Adversarial Fréchet Distance (AdvFD), which complements the static representation targets in FD-Loss with a calibrated adversarially learned representation. AdvFD augments the original static Fréchet objective with a learnable representation that adversarially maximizes the Fréchet divergence between real and generated samples, while the generator minimizes the same divergence in the resulting adaptive feature space. To prevent the adversarial representation from trivially increasing the objective through feature amplification, the authors further introduce real-feature whitening, which normalizes scale and covariance geometry and stabilizes the min-max optimization. Extensive experiments show that AdvFD consistently improves one-step generator post-training on both JiT and pMF backbones across different model scales.
Original Abstract
> 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. We attribute this failure to the static pretrained feature spaces used by existing Fréchet losses. These feature spaces provide incomplete and fixed views of the differences between real and generated distributions. To address this limitation, we propose Adversarial Fréchet Distance (AdvFD), which complements the static representation targets in FD-Loss with a calibrated adversarially learned representation…