Demosaicing Trace

A Bayer CFA is a mosaic of tiny color filters (RGGB) placed over a camera sensor. Each pixel captures only one color channel, and the camera's image processor interpolates the missing two — a process called demosaicing. Because an interpolated pixel is built from its neighbours, it sits closer to their average than a recorded pixel does, and that difference falls into a repeating 2×2 pattern. That is the trace this tool looks for. For broader metadata analysis, check your photo with the EXIF Checker or examine sensor noise with the Noise Analysis tool.

Each pixel is predicted from the mean of its four orthogonal neighbours and the leftover residual is squared into one of four buckets, according to its position in the 2×2 mosaic. The reported phase spread is (max − min) ÷ mean across those four buckets, and the threshold for calling the trace present is 0.25. Those numbers come from measurement, not from tuning: an image put through a real RGGB mosaic and a bilinear demosaic scores 2.019, while every sample photograph tested scores at or below 0.043. Unlike ELA, which reacts to recompression, this is a question about one specific step in a camera pipeline — and only that step.

The trace is fragile, and this is the limitation that matters more than any other. Measured on a controlled demosaic: 2.019 untouched, 1.527 after a JPEG q95 resave, 0.542 at q85, 0.388 at q75, and 0.015 after a single 0.75× resize — the last of these is indistinguishable from an image that never had a mosaic at all. Anything that has passed through a website, a messaging app, an export or an editor has almost certainly lost it. That is why this tool has no verdict and no red band: a result of not measurable on this file is the normal case, and it says nothing about whether the image was edited. Flat regions (sky, solid walls) leave nothing to predict from and show gray. Screenshots and rendered graphics never had a demosaicing stage. Read the result together with other forensic signals: check for clone regions, examine chromatic aberration consistency, or run the Authenticity Checker for a multi-signal verdict.

Not on its own, and it should not be used that way. It is true that an image from a diffusion model or a GAN never passed through a Bayer sensor and so cannot carry the trace. The problem is the other side of it: a genuine camera photograph that has been resaved or resized cannot carry it either, and every sample photograph tested here scored no higher than a screenshot. So a missing trace does not separate synthetic from real — it separates untouched-from-sensor from everything else. Use it as one input alongside the AI Detector and the FFT Spectrum Analyzer, and treat a found trace as informative while treating its absence as inconclusive. Explore all 131 forensic and utility tools at Scanly's full toolkit.