AdvFD: Boosting Visual Generation via Adversarial Fr'echet Distance Loss
AdvFD targets a failure mode where Fréchet scores improve while generated images do not.
The paper argues that fixed pretrained feature spaces can give generators an incomplete signal during post-training. Its proposed AdvFD adds an adversarially learned representation that searches for remaining distribution gaps while the generator tries to close them. The authors add real-feature whitening to stop the adversarial side from inflating the objective by scaling features. They report consistent gains for one-step generator post-training across JiT and pMF backbones and multiple model sizes. HF Daily Papers' note
The paper argues that fixed pretrained feature spaces can give generators an incomplete signal during post-training. Its proposed AdvFD adds an adversarially learned representation that searches for remaining distribution gaps while the generator tries to close them. The authors add real-feature whitening to stop the adversarial side from inflating the objective by scaling features. They report consistent gains for one-step generator post-training across JiT and pMF backbones and multiple model sizes. HF Daily Papers' note
score 5