NewThe detectors that scored perfect collapsed the hardest under attack.
Synthetic face, intermediate skin tone, femaleSyntheticIntermediateFemale
Examine the results

Check detectors from our latest benchmark

Pick a detector and a condition and see how it scores on our corpus, down to a single demographic cell. The same measurement we run in an evaluation. Full results are on our benchmark page.

Query the benchmark

Pick a selection and press Run.

AUC is a 0 to 1 score for how well a detector tells fake images from real ones. Picture showing it one fake photo and one real photo: AUC is how often it rates the fake as the more likely fake. 1.0 means it always gets it right, 0.5 is a coin flip (no better than guessing), and below 0.5 means it is getting them backwards.

Get involved

Put in a request

Public model

We'll run it and put it up on the interactive.

Private model

Share it privately. Weights or API Key.

Data and methodology

Self-service. Create a key and pull the data through the API.

Get an API key

Request a generator

Suggest one to add to the interactive.

AUC computed with tie-correct Mann-Whitney ranking, identical to our per-detector reports. Detectors are evaluated as published, with no fine-tuning.

Want the full report on your detector?

The interactive shows one score at a time. A full evaluation runs your model across the entire sequestered corpus and returns a per-group report card: AUC with DeLong confidence intervals, how it holds up under each platform condition, and where it fails, the recipe that broke it.