How to tell if an image was made with FLUX or Stable Diffusion
FLUX and Stable Diffusion are open models anyone can run on their own computer. No company stands behind them to attach a provenance signature, but the work record the generating tool leaves inside the file is a very strong clue.
The work record inside a PNG
Tools such as ComfyUI and AUTOMATIC1111 write the prompt, seed, sampler and node workflow into PNG text chunks by default. When that record is present the file is strongly shown to come from a generating tool without looking at pixels, and the report notes where it was and what it held.
When the record disappears
Re-saving as JPEG, posting to a messenger or social network, or passing through an editing app removes the text chunks. A missing record does not mean the image is not AI; it means the decision moves to the pixels.
What detectors see in open models
Open models come in many variants but share statistical regularities from the generation process. The box above shows the FLUX.2 family measured with our detectors. Fine-tuned derivatives and upscaled results can score differently.
The order to check in
Scan the PNG exactly as you received it so the text chunks are read first. If there is no record, read the detector result, and treat an inconclusive near the threshold as the conclusion.
Frequently asked
Is a PNG with the prompt removed still caught?
Once the record is gone the detectors decide. They catch most, but no guarantee covers every derivative model.
Does the report show the whole workflow?
The ledger records which chunks and keys were present, but the full prompt may be personal, so the screen shows only what is needed.
Is it the same for images made with a LoRA?
The base model's statistics usually remain, but results can differ. Read the result as a range.
Can I check locally only?
The Windows app for paid members (Pro or above) runs the same analysis without sending the file off the PC.