Know which model made the image.
We verify AI-generated images against the exact open-weight model they claim to come from. Base models, fine-tunes, even LoRA adapters. Not a guess from pixels: a measurement made with the model’s own weights.
How it works
You send the image and the shortlist.
A disputed upload, plus the models it might be. The claimed base, the suspected repackage, the fine-tune someone swears they trained themselves.
We test every candidate against the image.
Each model is scored on how well it explains the image it supposedly made. The model that made it fits. The others don’t. That’s the whole trick, and it only works because we hold the actual weights.
You get a verdict you can defend.
Which candidate fits, how confident we are, and what we can’t rule out. Every claim traces back to a measurement, with the error bars attached.
The pilot
Two weeks. Your hardest cases. Send us the disputed uploads, the suspected repackages, the license fights. In two weeks you get per-image verdicts, plus an honesty audit of whatever detection you already run. We test your current tools the same way we test our own instrument, including the failure modes most vendors never look for.
Keep the audit either way. If the numbers say our tool is wrong for your problem, you’ll hear it from us first.
Why trust us
We measure like it’s going to be checked. Every threshold we use was locked in before we saw the score it judges. When something in our own instrument broke, we found it, fixed it, and put it in the record. That happened four times, and the record is public to our clients.
So when we hand you a verdict, you get three things: the answer, the evidence, and a list of the things we refuse to conclude. Most vendors sell you the first one.
We also honor model creators’ license terms in everything we render. Provenance work starts with respecting provenance.
What we won’t tell you
Honesty about scope is part of the product, so here it is up front:
Whether an image is AI-generated at all. That’s a different problem, with different tools.
Anything about closed models. Midjourney, GPT-image and friends publish no weights. No weights, no measurement.
An answer when the true model isn’t on your list. We verify against candidates. We don’t divine unknowns, and we’ll say “inconclusive” before we say something we can’t back.
A number without its conditions. Accuracy figures live in our capability sheet, right next to the conditions that make them true. Ask for it.
Bring us your hardest case. If we can measure it, you’ll get a verdict you can stand behind. If we can’t, you’ll know exactly why. Either way, you’ll never get a confident answer we can’t prove.