Frequency-Domain Detection and Why It Traces to StyleGAN's Architecture

An interviewer asks: "Your image-level detector operates on a DCT (frequency-domain) representation of the face crop instead of raw RGB pixels. Why would that help, mechanically — what is it actually looking for?"

1. Answer the mechanical question: what does a GAN's synthesis network do that leaves a detectable frequency-domain signature, and why doesn't a real photograph show the same pattern?
2. A colleague argues "diffusion models don't use GAN-style upsampling, so frequency-domain detection is a GAN-specific trick that won't catch diffusion-generated fakes." Is this a complete argument? What would you want to check before accepting it?
3. Why does a detector trained only on frequency artefacts from still-image generators need continuous refreshing rather than being a one-time architecture decision?