Skip to main content

FreqCross — Detecting Stable Diffusion 3.5 Images

Author: Thomas
Original source: arXiv:2507.02995

Overview​

FreqCross fuses RGB, Fourier, and radial energy cues to expose images produced by Stable Diffusion 3.5. The three-branch network captures subtle spectral fingerprints that humans rarely notice, achieving 97.8% accuracy on a 10,000-image evaluation set.

Architecture in Brief​

  • Spatial stream: ResNet-18 backbone extracts classical vision features.
  • Frequency stream: Lightweight CNN analyzes FFT magnitude spectra.
  • Radial energy head: MLP inspects energy profiles to catch diffusion artifacts.

Visual Insight​

PCA of fused features (COCO vs SD3.5) on the test set

This PCA projection of the fused feature embeddings shows a clear separation between COCO (real) and SD3.5 (synthetic) samples on the test set, indicating that the frequency–spatial fusion captures discriminative artifacts beyond pixel-space cues.

Why It Matters​

  • Reveals distinctive frequency bands (0.1–0.4 normalized range) unique to diffusion models.
  • Provides a reproducible baseline with open-source code and models.
  • Offers content authenticity tools for media, platforms, and regulators.
Back to Research Portfolio