Paper Overview
- Field: Computer Vision (CV)
- Authors: Mingju Gao, Jingkai Zhou, Kun Gai, Changqian Yu, Hao Tang
- Published: 2026-08-11
- arXiv: 2608.11205
- A learnable representation adversarially maximizes the Fréchet discrepancy between real and generated samples.
- The generator minimizes the same discrepancy in this resulting adaptive feature space.
- To prevent the adversarial representation from simply inflating the objective through feature amplification, real feature whitening is introduced, normalizing the scale and covariance geometry of real features and stabilizing the min-max optimization.
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.
The authors attribute this failure to the static pretrained feature spaces used by existing Fréchet losses, which provide incomplete and fixed views of the differences between real and generated distributions.
To address this limitation, they propose Adversarial Fréchet Distance (AdvFD), which complements the static representation targets in FD-Loss with a calibrated adversarially learned representation:
Results
Extensive experiments demonstrate that AdvFD consistently improves one-step generator post-training across JiT and pMF backbones and different model scales.
---
*Auto-collected on 2026-08-13*