Reproduce “AdvFD: Boosting Visual Generation via Adversarial Fr'echet Distance Loss” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Mingju Gao, Jingkai Zhou, Kun Gai, Changqian Yu, Hao Tang
AdvFD enhances visual generation by integrating an adaptive, adversarially learned representation into Fréchet distance optimization, specifically mitigating the "Fréchet hacking" phenomenon. This approach consistently improved FID and held-out Fréchet distance metrics by up to 21.0% and 41.4% respectively, across various generative backbones on ImageNet-1K.
Key insights
- Relying solely on static feature representations for distribution-matching objectives can lead to "Fréchet hacking," where quantitative improvements do not correlate with qualitative visual enhancements or generalization.
- An adaptive, adversarially learned feature space provides a dynamic and evolving comparison metric that continuously exposes residual discrepancies, fostering more robust and perceptually aligned visual generation.
- Stabilization techniques like real-feature whitening are crucial for enabling stable and effective adversarial learning within feature spaces, preventing trivial solutions and maintaining the integrity of the comparison.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.