LUMEN / METHODS
What the detector examines and how to read the result.
Go to the check →Lumen accepts JPG, PNG and WebP up to 12 MB and 24 megapixels. Choose an image, drag it into the upload area or paste with Ctrl+V. Press “Check image”.
Use the original photograph where possible. Screenshots and copies from messaging apps alter details and may lose metadata. The shorter side must be at least 224 pixels for a full verdict.
The main detector is Community Forensics ViT-S/384, trained on real and generated images. Preprocessing preserves proportions. Besides the centre, the model examines two square regions at the top left and bottom right, each measuring 80% of the shorter side.
We check the original, a mirrored version and a JPEG copy at quality 85, then two offset regions. These are five passes of the same model, helping expose disagreement across the image. The 0–100 score shows its response to the AI class.
Community Forensics research ↗Borderline cases use Frontier, an additional model fine-tuned on newer generators. It checks the same five views separately.
Additional confirmation requires the three main variants to score at least 50 with a range of at most 20, and all five Frontier checks to score at least 98. In that case the headline score belongs to Frontier; both models’ scores are shown separately.
Signs of generation require all five scores to be at least 98. A photo verdict requires all five to be 10 or less. The large number is the lowest score. If the shorter side is below 224 pixels, the range exceeds 20 points or the thresholds are not met, the site explains the reason for uncertainty.
Retouching, computer graphics, heavy editing and unfamiliar generators can cause errors. Compare the result with the file’s origin and the additional measurements. A more complete original file often provides more information.
We look for software names and generation parameters in EXIF, PNG and XMP. Their absence alone does not identify the image’s origin: many services strip these records during upload.
For JPEG files, ELA maps differences after recompression at quality 90. It shows how image regions change. These measurements are separate and do not add points to the model score.
YuNet suggests a separate eye pair for each detected face, up to 20 pairs. Pairs are numbered by person and can be removed. For manual marking, mark both eyes of one person before the next. Confirm that both eyes are visible: the model can predict a closed eye’s position. Comparisons are made only within each pair.
We compare the selected regions’ brightness distribution using the Gini coefficient. You can also mark edges sharing a parallel direction and calculate their vanishing point.
Eye reflection research, RAS ↗The check works without registration. Photos are processed in server memory and are not stored by the application. You can download a JSON report after analysis.
Registration with a username and password keeps you signed in. The browser uses session and form-protection cookies. Details are on the data page.
Data and cookies →