Reproduce “Exploring The Effect of High-frequency Components in GANs Training” 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
Ziqiang Li, Pengfei Xia, Xue Rui, Bin Li
This research investigates how high-frequency components influence Generative Adversarial Network training, revealing the discriminator's over-reliance on these features. The study introduces simple frequency-domain operations, High-Frequency Filter (HFF) and High-Frequency Confusion (HFC), which significantly improve generated image quality across various GAN architectures and datasets by compelling the discriminator to focus on low-frequency content.
Key insights
- Discriminators in GANs are highly sensitive to high-frequency components and readily use them as shortcuts, even when these differences are imperceptible to human observers.
- This high-frequency bias in discriminator learning prevents the generator from effectively learning to replicate the semantically crucial low-frequency components of real images.
- By strategically manipulating the frequency components presented to the discriminator, its focus can be redirected to low-frequency content, thereby improving the generator's ability to produce more coherent and higher-quality images.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.