FLUX AI Image Detector

Add the picture to the FLUX AI image detector above. Images from Black Forest Labs' API carry Content Credentials signed by Black Forest Labs Inc., but that signer isn't on the C2PA trust lists our checker uses. So the checker treats the credential as a label it can't verify, lists it in the reasons and sends the picture to a detection model, which uses one of your free checks.

The model is Sightengine, with Hive as a backup if Sightengine fails, the same engines as the AI image detector on our home page. FLUX images made with open-weight models on someone's own computer carry no Black Forest Labs credential at all, so for those the model is the only signal from the start.

FLUX was the hardest family for both engines in our test. Sightengine caught 8 of 9 FLUX images and Hive 6 of 9, while Sightengine caught every image from the other five families. The checker got 8 right and 1 wrong, and that was its only wrong answer on any AI image in the whole test.

The miss was a screenshot. Sightengine scored it 0.18, below the checker's 0.2 line, so the checker called it Likely a real photo. The original file of the same picture was caught, as were the other screenshot and all 7 original files, FLUX.2 [pro] and FLUX.2 [max] alike. One miss in 9 isn't a rate you can rely on in either direction, but it is why we call FLUX the family to be most careful with.

Black Forest Labs signs its API images, and all 5 originals we generated through its API (via a reseller that relays requests to it) carried credentials saying they were AI-generated. The checker still doesn't let them settle the answer, because the signer isn't on the C2PA trust lists we check, and a credential from an unlisted signer proves less than it seems: anyone can sign a record like that with their own certificate.

Trust lists are how a checker knows a signer is who it claims to be. The C2PA launched its official trust list in mid-2025 and froze the older interim list on January 1, 2026; our checker accepts signers on either. Black Forest Labs' own help page says verification depends on the verifier's trust list, and that an untrusted-signer result means the verifier doesn't recognize the signer yet. In our test this cost nothing in accuracy, because the model called all 5 signed originals AI. If the signer reaches a trust list we use, files like these could be settled in your browser instead.

Two kinds of FLUX image reach the FLUX AI image detector with no credential at all: copies that lost it, and images from open-weight FLUX models run outside Black Forest Labs' service. For both, the detection model is the only signal.

The 2 FLUX.2 [pro] images we took from a public dataset had no credentials, and neither of our 2 screenshots kept them. Black Forest Labs says images made with self-hosted or open-weight models, such as FLUX.1 [dev] and FLUX.1 [schnell], are outside its signing service. We didn't test any FLUX.1 images or local runs, so we can't tell you how the engines do on them.

Because FLUX was the hardest family, give a FLUX picture a second look when the FLUX AI image detector surprises you. A Likely AI-generated answer was right every time in our test; the one miss went the other way, a FLUX screenshot called real.

• Get the original file if you can. Our only miss was a screenshot, and screenshots also lose the credentials.
• Read the reasons. If the credentials name Black Forest Labs API, that is a label worth noting, even though we can't verify the signer.
• Don't treat Likely a real photo as clearance on a picture that looks too polished or comes from an account that posts AI work. Use the visual checklist and look for the first post.
• Compare generators. The Midjourney and Stable Diffusion pages show families that were easier to catch, and the accuracy report has every result.

We can't say why; we can only report what happened. Both engines missed more FLUX images than any other family: Sightengine 1 of 9 and Hive 3 of 9. With 9 images, that is a warning sign, not a precise rate.

It reads them and shows them, but it doesn't let them decide, because Black Forest Labs Inc. isn't on the C2PA trust lists we check. The detection model decides, and the credential is listed as a label we can't verify.

No. It was a screenshot of a FLUX.2 [max] image, which Sightengine scored 0.18. The original file of the same picture was caught.

We didn't test FLUX.1 models or local runs, only FLUX.2 [pro] and FLUX.2 [max]. Sightengine lists Flux among the generators it aims to support, but that is the vendor's claim, not our measurement.