AdvFD: Boosting Visual Generation via Adversarial Fr'echet Distance Loss

TL;DR - AdvFD is a post-training loss for visual generators that pairs the standard Fréchet-distance objective with an adversarially learned feature space, addressing "Fréchet hacking" where the target metric improves while actual visual quality stagnates. It matters because it makes distribution-level generator post-training more robust than optimizing against fixed pretrained features.

• Diagnoses Fréchet hacking as a consequence of static pretrained feature spaces, which give incomplete and fixed views of the real-vs-generated distribution gap.
• Adds a learnable representation trained to adversarially maximize Fréchet discrepancy, while the generator minimizes it in that adaptive space — a min–max complement to the original static FD-Loss.
• Introduces real-feature whitening to normalize the adversarial representation's scale and covariance geometry, preventing trivial objective inflation via feature amplification and stabilizing training.
• Reports consistent gains for one-step generator post-training across JiT and pMF backbones and multiple model scales (no specific numbers given in the abstract).