Reproduce “Fourier Spectrum Discrepancies in Deep Network Generated Images” 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
Tarik Dzanic, Karan Shah, Freddie Witherden
Researchers identified systematic discrepancies in the high-frequency Fourier spectra of images created by deep generative networks compared to natural images. This enabled a low-data detection method, achieving 99.2% accuracy on uncompressed high-resolution images, and also informed a technique to modify generated images to mimic real image spectra.
Key insights
- Deep network generated images display systematically flatter decay rates in their high-frequency Fourier spectra compared to the characteristic power-law decay observed in natural images.
- These high-frequency spectral discrepancies are fundamental artifacts of deep generative models, enabling detection with high accuracy and minimal training data across various architectures.
- While common JPEG compression can homogenize the high-frequency spectra of some generated images, certain models like VQ-VAE2 produce artifacts more resilient to compression.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.