A model family is not a fingerprint
Black Forest Labs documents FLUX image generation and editing products, including hosted APIs, and some FLUX models are also run on people's own hardware through community tools. A hosted export, an edited image and a locally generated file can therefore follow very different paths. The model name alone is not a file fingerprint. We have not confirmed a universal FLUX provenance policy, so C2PA and SynthID coverage is treated as unverified here.
Source: Black Forest Labs documentation (opens in a new tab)
Start with provenance and metadata
Keep the earliest available original and note where it was downloaded. Local tools sometimes write generation details into the file: ComfyUI, for example, can embed the full workflow in saved PNG images so the file can be reopened to recreate the setup, and other interfaces save prompts and settings as PNG text. When present, those fields are strong clues worth recording. When absent, they say little, because metadata is easy to strip.
Source: ComfyUI examples: workflow metadata in images (opens in a new tab)
Do not overread realism
Photographic-looking textures, convincing shadows or readable text do not prove camera capture. Image-generation tools can produce plausible scenes, while real photographs can contain strange exposure, perspective or compression artifacts. Instead of assembling a checklist of aesthetic flaws, examine the claim attached to the image. Ask whether there is an original capture, an independent source or a documented editing process that supports the specific event being asserted.
Understand model coverage
A detector can fail on FLUX versions and workflows that were missing from its evaluation set, and fine-tuned or locally modified models can shift results further. A meaningful test uses original exports and transformed copies with known labels, and it reports false positives on camera images alongside detection rates. Until such measurements exist, Image Evidence makes no FLUX accuracy, recall or generator-attribution promise.
Use the result carefully
A metadata finding, a model estimate and an external provider's assessment are different evidence types. Agreement between them can guide further review but is not a legal certification. Write down what you know about a file's origin and what remains unknown, and avoid presenting a single score as proof. Image Evidence does not train on user content.
Can you detect FLUX-generated images?
Sometimes, with caveats. Embedded generation metadata can be decisive when it survives, and some general detectors flag FLUX images, but no public tool reliably identifies every FLUX version, especially after editing, resizing or compression. Look for published evaluations that include FLUX outputs and real photographs before trusting a score.
Do FLUX images contain metadata?
It depends on how they were made. Files from local workflows may carry prompts, seeds or full workflows in PNG text chunks, while hosted services and apps may write different metadata or none at all. Social platforms often re-encode uploads and drop these fields. A metadata viewer such as ExifTool can show what remains in a copy of the original file.
Source: ExifTool (opens in a new tab)
Is a FLUX image with no metadata a real photo?
No. Missing metadata is the normal state of many shared images, including AI-generated ones. Treat it as an absence of evidence and move on to the source, the claim attached to the image and any available provenance records.