How to tell if an image is Flux 2
FLUX.2 is the image model family Black Forest Labs released on 25 November 2025, in [pro], [max], [flex], [dev] and [klein] variants. It resolves most of the artefacts that used to give AI images away, so identifying it visually is harder than it was with FLUX.1, and the reliable signals have moved into the file's metadata. What provenance an image carries depends almost entirely on how it was generated: images made through the Black Forest Labs API are signed, images made from the open weights on someone's own machine usually are not.
Visual tells
The old checklist no longer works. Counting fingers, looking for melted ears or nonsense lettering will not clear a FLUX.2 image — the model handles hands, anatomy and short text correctly most of the time. Treat a clean-looking image as unresolved, not as verified real.
Text degrades with length, not with difficulty. Short signage, product labels and single words usually render correctly. Errors concentrate in paragraph-length text, dense small type, and repeated labels across a scene, and they tend to show up as invented extra words or a string that runs on past what the scene needs.
Resolution tops out around 4 megapixels. Black Forest Labs documents FLUX.2 output and editing at up to 4MP. A photographic-looking image capping out near 4MP is inconsistent with a modern camera, which typically produces 12MP to 50MP files.
Dimensions divide evenly by 16. The open-weight implementation builds its latent at height/16 by width/16, so output sides are multiples of 16, and the reference tool's default is 1360x768. Camera dimensions like 4032x3024 or 6000x4000 come from sensor geometry instead, so an odd 16-divisible size such as 1360x768 is worth noting.
Composite seams in multi-reference work. FLUX.2's main feature is building an image from several reference images — the pro API endpoint accepts eight reference slots. Where an image was assembled this way, check whether the lighting direction, colour temperature and grain on a subject actually agree with the scene behind it.
Scenes that are more specified than the situation warrants. FLUX.2 [pro] and [max] apply prompt upsampling by default, expanding a short prompt into a detailed one. The result is often a scene where every object is deliberately placed and lit, with none of the incidental clutter, dead space or awkward framing of an unstaged photograph.
Metadata & provenance
Provenance depends on the route, and the difference is large. Through the Black Forest Labs API: the FLUX.2 [pro] and [klein] model cards state that the API "applies cryptographically-signed C2PA metadata to output content to indicate that images were produced with our model." So an unaltered API image should carry a verifiable C2PA manifest. The API's default output format is JPEG, with PNG and WebP available. From the open weights, run locally: no C2PA is written. The published inference repository does ship a pixel-layer invisible watermark (the imwatermark library, dwtDct mode, carrying a fixed 48-bit message), but in the official command-line tool both the import and the call to it are commented out — so it is off unless someone deliberately switches it on. Black Forest Labs describes this as "an example of pixel-layer watermarking" and separately recommends that deployers implement C2PA themselves. In practice, assume no invisible watermark on locally generated FLUX.2 images. EXIF: the official command-line tool does write two tags — Software set to "AI generated;flux2" and Make set to "Black Forest Labs". This is worth checking because it is an explicit, unambiguous marker. But it only appears if the image came from that specific tool. Images generated through ComfyUI, diffusers or a third-party host will not have it. ComfyUI, which is how many people run the open weights, instead writes its full workflow and prompt into PNG text chunks, which will name the FLUX.2 checkpoint. No SynthID. SynthID is Google DeepMind's system, used across Google's own models and licensed to some other vendors. Black Forest Labs has not adopted it, and it is not part of the FLUX.2 release. Two caveats that matter more than any of the above. First, all of this metadata is removable: EXIF comes off with any image editor, and C2PA manifests are stripped as a routine side effect of the recompression that Instagram, X, Facebook, LinkedIn and TikTok apply on upload. An image that reached you via social media has almost certainly lost its credentials, and their absence proves nothing. Second, the reverse also holds — the presence of a valid C2PA manifest is evidence, but its absence is not evidence of anything at all.
How well WeDetect handles Flux 2
We have not published a FLUX.2-specific accuracy figure, and we are not going to imply one. Our published numbers are roughly 89% overall across modern generators, and about 85% recall on Midjourney, DALL-E, SDXL and FLUX.1. That FLUX.1 figure should not be read as a FLUX.2 figure: FLUX.2 is a different and newer model, and detector performance is known not to transfer cleanly across generator versions. For context on how much this varies, our measured recall on Stable Diffusion 3.5 is about 62% — meaning we miss roughly four in ten SD 3.5 images. We would rather state that plainly than let readers assume a uniform number. Every result we return is probabilistic. A "likely AI" verdict is not proof, a "likely real" verdict is not a clearance, and neither should be the sole basis for a consequential decision. Where an image still has its metadata, the metadata checker gives a firmer answer than the pixel model does — a valid C2PA manifest naming Black Forest Labs, or an EXIF Software tag reading "AI generated;flux2", is much stronger evidence than any classifier score. Where metadata has been stripped, the pixel model is all that is left, and it can be wrong in both directions.
Check an image now
Drop the image in below — free, no sign-up, deleted after analysis.
Drag & drop, paste, or click to upload
PNG · JPG · WEBP · HEIC — your image is analyzed privately and deleted
This is an estimate, not proof — the detector can be wrong, producing both false positives and false negatives. Don't rely on it as your only source; verify anything important. wedetect.ai is not liable for decisions made from these results. See our Terms.
Frequently asked questions
Does Flux 2 add a watermark to images?
Not by default in the version most people use. The published inference code includes an optional invisible pixel watermark, but it is commented out in the official command-line tool, so images generated from the open FLUX.2 weights normally carry no invisible watermark. There is no visible watermark either, and Flux does not use Google's SynthID. Images generated through the Black Forest Labs API get signed C2PA metadata instead, which is a different mechanism and is easily stripped.
Do Flux 2 images have C2PA Content Credentials?
Only if they came through the Black Forest Labs API. Black Forest Labs states that its FLUX.2 [pro] and [klein] API applies cryptographically-signed C2PA metadata to output. Images generated locally from the open FLUX.2 [dev] or [klein] weights get none. And because Instagram, X, Facebook, LinkedIn and TikTok all recompress uploads, a manifest that existed at generation time is usually gone by the time you see the image. Finding a valid manifest tells you something; not finding one tells you nothing.
How can I tell if an image is Flux 2 or Flux 1?
Usually you cannot, from the pixels alone. Neither we nor any detector we know of reliably distinguishes FLUX.2 output from FLUX.1 output by appearance. Metadata is the only dependable route: a C2PA manifest or an EXIF Software tag reading "AI generated;flux2" names the version outright, and a ComfyUI PNG will have the checkpoint name in its embedded workflow. Failing that, weak circumstantial hints are that FLUX.2 handles short text and hands noticeably better, and that FLUX.2 output can reach 4MP where FLUX.1 was typically generated smaller.
Can Flux 2 images be detected?
Sometimes, but not dependably enough to treat a single result as proof. FLUX.2 is one of the harder cases: it produces few of the artefacts that classifiers were built around, and the open weights usually ship with no metadata to check. Our detector is probabilistic and will miss some FLUX.2 images and occasionally flag real photographs. For anything with consequences attached, use it alongside other checks — reverse image search, the provenance of the file, and whether the person supplying it can produce the original with its camera metadata intact.
Other generators: Midjourney · DALL·E 3 · GPT-4o · Stable Diffusion · Flux · Grok · Firefly · Gemini · SDXL · Nano Banana · GPT Image 2 · SD 3.5 · Sora · all generators.