How to Tell if an Image Was Made With FLUX
FLUX is not one tool. It is a family of models you can rent through a hosted API or run on your own hardware, and that split decides whether an image carries any provenance at all. Hosted services can sign output with Content Credentials. A self-hosted copy usually will not, because the labeling code is optional. So check the file first, then fall back to the pixels.
Hosted APIs and open weights behave differently
The hosted versions of FLUX and Stability models can attach C2PA metadata at the service layer, which gives you a signed record of how the image was made. The open weights do not require it. The official FLUX.2 repository includes watermarking and provenance code, but a self-hoster can simply leave it out, and many do. That is the structural problem with marking open models: any signal baked into open inference code is optional by design.
Check the file before the pixels
Inspect the original file for Content Credentials at contentcredentials.org/verify. If a manifest names a hosted FLUX service and validates, that is real evidence. If the file has no manifest, you have learned almost nothing, because both a stripped hosted image and a self-hosted image look the same. A screenshot removes the manifest too, so ask for the original when the answer matters.
What FLUX images tend to look like
- Photoreal. FLUX.1.1 Pro earned its reputation on realism and close prompt following, and FLUX.2 continues that. Skin, fabric, and glass often look convincing, so lighting mismatches are a better clue than surface detail.
- Hands and text. Complex hands, jewelry, and small or curved text still fail more often than the rest of the frame.
- Background logic. Follow fences, railings, and window frames. Repetition and impossible architecture survive across models.
The full method is in our guide to spotting AI-generated images.
Use a detector as one signal
Score the file with the free AI image checker or the app and read the result as a probability. FLUX output is good enough to land in a mid-range score, and a single number will not settle it. We cover the failure modes in Do AI Image Detectors Actually Work?. For a neighboring model, see the Stable Diffusion guide.
Ask for the original
When the image matters, request the original file, the prompt, and the workflow. A creator who ran the model themselves can supply them, and that comparison is stronger than any detector score. It is also the only check that still works when every mark has been stripped.
Common questions
Does FLUX add C2PA Content Credentials?
It depends on where the image was made. Hosted FLUX services can sign images with C2PA at the service layer. Open-weight FLUX, run on your own machine, ships labeling code that a self-hoster can remove, so the same model can produce marked and unmarked images.
Is FLUX watermarked?
The official FLUX.2 repository includes watermarking and provenance code, but it is optional in a self-hosted setup. Treat a FLUX image as unmarked unless you can read a C2PA manifest from the file.
How do I check a FLUX image, then?
First inspect the original file for Content Credentials. If there are none, fall back to the visual checks and a detector score, and treat both as signals rather than proof.
Which FLUX model is current?
FLUX.2 is the current family from Black Forest Labs. Earlier versions such as FLUX.1.1 Pro built its reputation on photorealism and close prompt following, which is why FLUX output often passes a quick look.
Score a suspect image in seconds
AI Detector checks the image and never stores it.
