I8Image forensicsGeneric ForensicsLayer 1 (Deterministic)

Edge Coherence

Checks whether the sharpness of edges is consistent across the entire image. Spliced or AI-generated regions often have edges that are sharper or softer than the surrounding content.

Technical description

Measures how uniform the edge sharpness is across the image. The Sobel edge magnitude G = sqrt(Gx^2 + Gy^2) is computed, the image is tiled into an 8x8 grid, and each block's edge density is the mean of G over the block divided by 255. With per-block densities of mean d_bar and standard deviation s, the coefficient of variation CV = s / d_bar is taken. The cue is gated on content (mean edge density above 0.03), so flat images are never judged; when there is content and CV is below 0.35, the contribution is (0.35 - CV)/0.35 * 3.5, capped at 3.5 because this is a weak supporting signal. A natural scene varies its sharpness through depth of field and texture and has a high CV; a frame rendered at one uniform sharpness has a low CV.

How it works

Layer 1 (deterministic). Computes the Sobel edge magnitude, the per-block edge densities on an 8x8 grid, and their coefficient of variation. Flags an unnaturally uniform distribution (low CV) only when the image carries enough edge content, capping the contribution as a weak cue. Reports the mean edge density, the edge-density CV, whether there was content to judge, and the block count.

Why this matters

How sharpness is distributed across an image is a recognised forensic cue. Blur and sharpness inconsistency is a classic basis for splicing localisation, since an inserted region carries its own focus and blur that differs from the host. Synthetic images also differ from real ones in local structure, and media-forensics overviews place local sharpness and texture among the distinguishing cues. A real lens and scene rarely render the whole frame at one sharpness, so an unnaturally uniform edge distribution is a soft sign of synthetic origin, used alongside the stronger screens.

Score thresholds

0-1: Edge sharpness varies naturally, or the image is too flat to judge
2-3: The edge sharpness is noticeably uniform across a content-rich image
4-5: A strikingly uniform sharpness across the frame, a soft cue toward synthetic origin

Limitations

A deliberately weak signal with clear confounders. A uniform-texture photograph (grass, fabric, gravel, foliage) fills the frame with even edge density and can read as uniform without being synthetic, and a uniformly defocused or motion-blurred photo is uniform for an honest reason; conversely a generated image with simulated depth-of-field varies its sharpness and passes. The content gate prevents flat-image false positives but leaves them unjudged. The 8x8 grid and global CV give a coarse whole-image view that does not localise. Splicing by compression history, noise consistency, copy-move, and the stronger frequency and colour cues live in sibling indicators.