FLUX image detector

Why FLUX images are hard to spot by eye

When FLUX.1 arrived in August 2024 it closed several gaps at once: legible text in signs and labels, correct hands, convincing skin and realistic phone-camera framing. Prompts like “amateur photo, iPhone, 2015” produce pictures that look like ordinary snapshots rather than glossy renders. Because the [dev] and [schnell] weights are downloadable, the model was quickly built into hundreds of apps, bots and fine-tunes, often unlabelled.

Some visual habits remain, and people who look at many FLUX images learn them: a characteristic cleft chin on faces, very smooth backgrounds with heavy bokeh, plastic-looking skin in close-ups, and the same few faces recurring across unrelated prompts. They are hints, not tests — fine-tunes and LoRAs remove them.

What GPTTrace checks on a suspected FLUX image

Workflow metadata

People running FLUX locally mostly use ComfyUI or Forge, which save their settings into the PNG. A ComfyUI workflow loading a flux1-dev checkpoint, or a parameters chunk naming a FLUX model, is shown as proof. GPTTrace decompresses compressed chunks too.

Resolution

FLUX works on a latent grid of 16-pixel patches and is commonly run at about one megapixel — 1024×1024, 1344×768, 832×1216 and similar. A camera photo rarely has such dimensions.

Pixel statistics

FLUX uses a 16-channel autoencoder rather than the 4-channel one of earlier Stable Diffusion models, so some older forensic tricks no longer apply. The neural classifier still finds generation traces in most FLUX output; you will see its score and the per-crop scores in the result. Its confidence drops on screenshots, heavy JPEG compression and small images under about 400 pixels.

FLUX editing models change the question

FLUX Kontext and later editing models do not only create images from scratch; they take a real photo and change it — swap the background, alter a sign, put a different expression on a face — while keeping everything else pixel-for-pixel similar. That produces a hybrid: most of the image came from a camera, a region came from a model. A whole-image classifier may give such an edit a middling score. If the result is inconclusive but something in the picture is the point of the story (a sign, a face, an object in someone’s hand), crop to that region and check the crop on its own: the classifier will see a larger share of generated pixels.

Where FLUX images show up

Because it is free to run and good at ordinary-looking photos, FLUX is a common source of fake profile pictures, fabricated “eyewitness” shots and product images for listings that don’t exist. If you are checking a person rather than a picture — a dating match, a new LinkedIn contact — see the fake profile picture checker for the non-technical checks that work alongside this tool.