Mingju Gao, Jingkai Zhou, Kun Gai, Changqian Yu, Hao Tang
4 min
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. AdvFD augments the original static Fréchet objective with a learnable representation that adversarially maximizes the Fréchet discrepancy between real and generated samples, while the generator minimizes the same discrepancy in the resulting adaptive feature space. To prevent the adversarial representation from trivially increasing the objective through feature amplification, we further introduce real-feature whitening, which normalizes its scale and covariance geometry and stabilizes the min--max optimization. Extensive experiments show that AdvFD consistently improves one-step generator post-training across both JiT and pMF backbones and across different model scales.
Sam: So it's not just that the target moves—it's that the model can't cheat by leaning heavily on one dimension. It has to improve across the board.
Alex: Precisely. And that combination is what makes the approach stable. The adversarial part keeps the model from settling into any one shortcut. The whitening part prevents it from finding a new shortcut in a different direction. Together, they push the model toward output that genuinely resembles real images, rather than output that merely tricks the scoring formula.
Sam: So the insight is really about the nature of the evaluation itself. A rigid, static measure will always be vulnerable to being gamed. A flexible, adaptive one is much harder to fool.
Alex: That's it. And the study suggests this holds up across models of different sizes, which points to it being a general principle rather than a fix that only works in one narrow case. When the evaluation process is more honest, the images that come out of it are more genuinely realistic. It's a reminder that in machine learning, how you measure progress matters just as much as what you're trying to build. Thanks for listening to ResearchPod.