LUMEN / PHOTO FORENSICS
Check provenance, model scores and image details.
Check a photo →Keep the original file and the publication link. Use reverse image search to look for earlier copies. Compare the date, caption and other photos of the same scene. A messaging-app copy usually contains less information than the original.
Upload the file to Lumen and open the model scores. The centre, mirrored version, JPEG copy and two offset regions are checked separately. Disagreement helps show which region changes the answer. Rechecking the same file does not add new evidence.
For a camera photo, compare the original with the published copy. Retouching, skin smoothing, denoising and recompression change features learned by the model. Inspect image dimensions, metadata and score spread. Do not manually adjust a verdict just to fit one example.
Compare each person’s own two eyes. Remove a proposed pair if an eye is closed, covered by hair or too small. Differences in brightness can depend on lighting and the selected box. Lumen’s Gini comparison measures region brightness; it does not isolate reflections as the RAS experiment does.
New generators can produce details missing from training data. Lumen uses Frontier in borderline cases. The models share training ancestry, so their agreement is not equivalent to two independent investigations. Look for additional context: the original publication, a sequence of photos and software records in metadata.
Download the JSON report to keep variant scores and measurements. Record where the file came from and whether it was edited. Use matching image versions for meaningful comparisons. The Lumen methodology describes the current thresholds and file preparation.