Image Watermark Detector

The Detector runs three pixel-based checks (visible overlay, LSB statistic, and the invisible frequency pattern with a blind text reading) and parses the file's metadata for copyright/rights fields and AI-provenance declarations. The table lists these as five rows, and Standard mode lets you switch four categories on or off (the two metadata rows share one). It is built around the marks the Image Watermark Embedder writes, plus the declarations an image file can carry; it is not a general-purpose detector for every watermark product. This detector does not detect or verify SynthID or other invisible AI watermark systems. It reads selected provenance declarations stored in image metadata, but it does not establish whether an image was AI-generated. The declarations it can read are a C2PA / Content Credentials manifest and the IPTC or XMP DigitalSourceType field; they may be lost when an image is re-saved, screenshotted or stripped of metadata.

No detection result establishes whether you have permission to use an image. A LOW signal level means only that these checks found nothing; it does not mean the image is free to use, because copyright generally exists without any watermark.

Visible overlay: the image is divided into five regions (the four corners and the centre, up to 100 px square). For each, the tool counts adjacent pixels that differ strongly and pixels that are not fully opaque, and compares that with the same measure over the whole image. A region clearly above the image's own level is reported, with its position. Comparing with the whole image is what stops a uniformly textured photograph from being flagged. It is informational only: it is shown in the table but adds nothing to the score and never changes the signal level, because it cannot reliably tell an overlay from a busy area. We measured it on 37 clean real photographs (7 full-size camera photographs of 3 to 16 megapixels, 14 crops of them at 800 × 600, and 16 crops of one wallpaper photograph): it flagged 8 of them. With the Embedder's default text overlay (white, bold, 28 px, 40% opacity, bottom-right) added to the same 37 images it detected 28 and named the right position in 23; the result was identical after a JPEG re-save at quality 95 and at 75. Across a sweep of 864 overlays (six positions, three opacities, three sizes, white and black text, on eight photographs) it detected 609, but that count includes the photographs it flags without any overlay, and for the tiled overlay it never named the right position (0 of 144). We also tried a redesigned check based on the fact that an alpha-blended single colour lifts every stroke pixel by the same share of its distance to that colour: it separated an overlay clearly on smooth backgrounds, but on busy photographs the figures of the overlay and of clean regions overlapped, so it was not adopted. Treat the row as a pointer to look at the image, not as evidence of a watermark.

LSB statistic: for each of R, G and B, a chi-square test (Westfeld and Pfitzmann) asks whether the counts of the two values in each pair 2k and 2k+1 are unusually equal, which happens when the lowest bits were overwritten with random-looking data. Smooth gradients and flat graphics also have equal pairs, so a channel counts only when its lowest-bit plane also looks random at distances 1 and 2 in both directions. The statistic only rises when almost the whole lowest-bit plane was overwritten: in our tests a full random overwrite was flagged in 16 of 16 photographs, 25–75% overwrites in none, and clean photographs in none.

Invisible frequency pattern: the brightness channel is transformed (Haar wavelet, then an 8 × 8 DCT of the LL subband) and, for each of the four embedding strengths, the share of blocks whose largest coefficient sits near a quantization target is scored. A score of 0.75 or more counts as detected. Separately the Detector tries to read the text: for every strength and every text length from 4 to 128 bytes it takes the majority vote of the repeated bits and accepts a reading only when the copies agree, the text is readable, at least 12 copies of each bit exist and the reading is consistent at twice the length. A text read this way is strong evidence in itself, so it also counts as a detection. The strength shown is approximate: neighbouring strengths can be confused.

Metadata: EXIF Copyright and Artist (IFD0), IPTC Copyright Notice and By-line, XMP rights, creator and credit, and the PNG text fields Copyright and Author are read from JPEG, PNG and WebP files (the parser handles these layouts; it was tested on files built to them, not on files from every camera or program). A copyright or rights field raises the level to HIGH; authorship alone (an Artist or Author name) raises it to MEDIUM.

HIGH: the frequency pattern (score at least 0.75, or a text read blind), or a copyright or rights field in the metadata. MEDIUM: an LSB statistic, or authorship-only metadata (an Artist or Author name). LOW: nothing found. The LSB statistic is weak evidence that also appears in some clean images, so on its own it only reaches MEDIUM. The visible-overlay heuristic is shown for information but is not counted: it adds nothing to the score and never changes the level. The level is a summary of evidence, not a probability and not a legal finding; the level itself is never a percentage. Each row that fired shows its own strength on a 0–100% scale, which is not a probability.

A watermark you embedded is not found: check whether the image was resized, cropped, rotated, screenshotted or printed afterwards (in our tests a copy kept at a smaller size and small crops lost the mark every time; other edits were not tested systematically; see the table above). Check also that the text is not too long for the image: the Detector attempts a blind reading only with at least 12 copies of each bit (a rule built into it; longest readable text: 19 bytes in 800 × 600, 4 in 400 × 300). A pattern can still be flagged without readable text.

A clean image shows a flagged row: the visible-overlay row (informational, it does not change the level) can fire on a locally busy region, and the LSB statistic on an image whose lowest bits were genuinely rewritten (an editor, a converter, added grain). Look at which row fired before drawing a conclusion; a frequency pattern or a rights field is much stronger evidence than either.