Scans the first 64KB of file bytes for AI software signatures: Stable Diffusion, Midjourney, DALL·E, Grok, ComfyUI, Flux, Kling, Runway, and 20+ others. For JPEG files, checks for the presence of an EXIF APP1 segment — missing EXIF in a high-resolution JPEG is a significant AI indicator. GPS coordinates and camera make/model are treated as authenticity signals (reduce score). PNG iCCP and tEXt chunks containing parameters, steps, or cfg scale fields are direct Stable Diffusion fingerprints.

Weight auto-reduces to 0% for videos (metadata stripped by codec) and screenshots (no EXIF from screen capture).

Divides the image into non-overlapping 8×8 pixel patches and computes luminance variance per patch. Diffusion models produce unnaturally smooth mid-frequency regions — patches with variance < 15 are flagged. A smooth ratio above 45% of all patches triggers a high-confidence AI signal.

Also measures RGB inter-channel balance. AI images tend to have artificially equal R/G/B channel distributions (average channel difference < 18) because they synthesise colour globally rather than inheriting sensor noise. A variance-of-variance metric checks whether patch variance is itself too uniform across the image — another diffusion hallmark.

Applies a row-wise 1D Fast Fourier Transform to a 128×128 centre crop of the image. Computes three ratios: DC energy concentration (very smooth image), periodic peak energy at quarter-frequency (GAN upsampling fingerprint), and high-frequency energy ratio (over-smoothing signal).

This layer's weight has been deliberately reduced from 20% to 8%. GAN-based generators (StyleGAN, BigGAN) leave strong periodic spectral peaks, but modern diffusion models (Stable Diffusion XL, Flux, Grok) do not. Keeping the weight high causes false confidence for diffusion images. It remains useful as a tiebreaker for older GAN-based synthetic media.

Loads a 224×224 canvas slice and converts it to a normalised TensorFlow.js tensor in the range [−1, 1]. Computes four statistical signals without any trained classification model:

• Global mean — diffusion outputs are over-normalised, producing means very close to 0.0 (|mean| < 0.05 → strong signal)
• Inter-channel variance balance — AI images have suspiciously equal R/G/B variances (< 0.015 difference → strong signal)
• Kurtosis — diffusion images produce flat pixel distributions (kurtosis < 2.2 is highly indicative; real photos typically show kurtosis > 3.0)
• Local contrast entropy — divides image into 8×8 blocks of 28px; low mean contrast (< 60) combined with low block std dev (< 20) indicates artificially uniform local detail

This approach requires no labelled training data and generalises well across generator architectures. TF.js is loaded lazily — only when detection begins — and the tensor is disposed immediately after stats are extracted.

For JPEG files: computes bytes-per-pixel ratio (unusually small → over-processed AI image) and scans DCT block boundaries at every 8th row for blocking artifacts. Counterintuitively, very clean DCT boundaries (ratio < 0.02) in a JPEG also indicate AI generation, because diffusion models produce images without the block-boundary errors of real-camera encoding.

For PNG files: checks whether dimensions match common AI output sizes (512, 768, 1024, 1536, 2048px) or are exact multiples of 64 — the diffusion diffusion model latent space tile size. Also checks for exact standard aspect ratios (16:9, 4:3, 1:1 etc.) which are far more common in AI outputs than in real photography.

Re-encodes the canvas as JPEG at quality 75 using canvas.toDataURL('image/jpeg', 0.75), reloads the output as an image, and computes the absolute per-pixel difference between original and re-encoded versions.

The key metric is the Coefficient of Variation (CoV) of the error distribution: CoV = std(errors) / mean(errors). Authentic photographs show non-uniform error maps — edges and high-detail areas have higher re-encoding error than smooth regions. AI diffusion images are generated at globally consistent quality, producing unnaturally uniform error maps (CoV < 0.5 = strong signal; CoV < 0.8 = moderate signal).

Additionally, patch-level ELA uniformity is checked: the image is divided into 8×8 coarse patches and the standard deviation of patch-level mean errors is computed. Real photos show high variation between patches; AI images show minimal patch-to-patch error variation.

Computes the average luminance variance across non-overlapping patches at three scales: 4px, 16px, and 32px. In authentic photographs, variance is self-similar across scales — a property arising from the fractal nature of natural scenes. The 16px/4px variance ratio in real photos typically falls between 0.7 and 1.3.

AI diffusion models over-smooth at medium and coarse scales, producing a sharp variance drop-off: 16px/4px ratio < 0.35 contributes 40 score points. A 32px/16px ratio < 0.4 contributes 30 additional points. This layer is effective even after JPEG recompression because the multi-scale structure is a property of image content, not encoding artifacts.