New Paper Proposes AdvFD: Using Adversarial Training to Prevent FID Hacking in Visual Generation

kwangmoo_yi · x · 2026-08-13

The reviewer highlights a new paper by Gao and Zhou et al. titled AdvFD. While current visual generation models often use FID (Fréchet Inception Distance) as a loss function, this can lead to models 'hacking' the metric. The paper proposes an Adversarial Fréchet Distance loss, utilizing an adversarial training setup to prevent this exploit.

Related event: AdvFD Tackles Fréchet Hacking in Visual Generation Models(3 posts)→

Original post →

More from Research

Research channel →