Preventing Metric Hacking in Visual Generation: The AdvFD Loss Paper

kwangmoo_yi · x · 2026-08-13

Directly using Fréchet Distance (FD) as a loss function for generator post-training can lead to Fréchet Hacking: the target metric keeps improving while the actual visual quality stagnates or degrades. This happens because existing FD losses rely on static, pretrained feature spaces.

To address this, a joint paper by Peking University and KlingAI Research proposes AdvFD (Adversarial Fréchet Distance). The method introduces a learnable representation that adversarially maximizes the Fréchet discrepancy while the generator minimizes it, creating an adaptive min-max optimization. To prevent the representation from inflating the objective via feature amplification, the authors add real-feature whitening to stabilize training. Experiments show consistent improvements in one-step generator post-training across backbones like JiT and pMF.

Original post →

Preventing Metric Hacking in Visual Generation: The AdvFD Loss Paper

More from Multimodal