Image Detection Concepts

AI-generated images often leave behind subtle artifacts that are invisible to the naked eye but detectable through analysis.

Frequency Analysis

Generative models (GANs, Diffusion) often introduce artifacts in the frequency domain (e.g., checkerboard patterns) due to upsampling operations. FrequencySignal uses FFT (Fast Fourier Transform) to detect these anomalies.

Reconstruction Error (DIRE)

DIRE (Diffusion Reconstruction Error) relies on the idea that diffusion models can reconstruct images they generated better than real images. By passing an image through a diffusion autoencoder and measuring the reconstruction error, we can infer if it was generated by that model.

Semantic Inconsistency (CLIP)

CLIPSignal uses zero-shot classification to check if an image matches "AI generated" prompts better than "real photo" prompts in the CLIP embedding space.

Mean Local Entropy (MLEP)

MLEP (Mean Local Entropy Pixel) analyzes the texture details of an image. Synthetic images, especially those from earlier GANs or unrefined diffusion models, often exhibit different local entropy characteristics compared to natural images. High local entropy might indicate noise or artifacts, while unnaturally low entropy might indicate "smoothing" or lack of detail.