Pixel Forensics
How It Works
We analyze pixel-level patterns to detect potential synthetic characteristics:
π FFT Frequency Pattern
Measures high/mid/low frequency content distribution. Real photos typically have more high-frequency details (edges, texture, noise).
Baseline: Real median 0.472, AI median 0.268 (AUC 0.925)
π Noise Residual
Analyzes sensor noise patterns. Real cameras introduce characteristic noise; AI images tend to be smoother.
π Neighbor Pixel Correlation
Measures local pixel relationships. AI generation may show subtle correlation artifacts.
π Frequency High/Low Ratio
Compares high-frequency to low-frequency energy. Strong discriminator between real and synthetic images.
β οΈ These Are Auxiliary Signals
- Pixel forensics are not proof of AI generation
- They serve as statistical indicators, not definitive tests
- Compression, cropping, and editing affect these signals
- Some modern AI tools produce near-natural pixel patterns
- Always combine with metadata evidence for reliable analysis
β When Pixel Forensics Help
- Confirming AI evidence when metadata is present
- Detecting synthetic patterns in metadata-stripped images
- Identifying frequency anomalies as weak indicators
- Supporting human judgment in ambiguous cases