Open-weight image models and provenance: why local generations often carry no label

By The NeutronField editorsPublished Updated

Short answer

Open-weight models such as Stable Diffusion and Flux run on your own machine, so no company signs the output. Their licences mostly leave labelling to the law, invisible watermarks sit in optional reference code that front ends can switch off, and no content credential is added locally. What local tools usually do save is your prompt and settings, as plain text in the file.

What open-weight means

Open-weight means you can download the trained model and run it yourself, under the maker's licence, which is not always an open-source one. Stable Diffusion 3.5 Large uses the Stability AI Community License, free for organisations or individuals "with less than $1M in total annual revenue" (Stability AI). FLUX.1 [schnell] is released under Apache 2.0 (Black Forest Labs), while FLUX.1 [dev] and FLUX.2 [dev] use non-commercial licences for the weights (Black Forest Labs).

frame 64An open padlock and a ring of old keys on a workbench in a home workshop
▸ 64NF 08
Frame 64. Open weights hand you the keys; labelling is up to you.
The FLUX.1 [dev] licence still lets you use the images themselves "for any purpose (including for commercial purposes)" unless the licence says otherwise (FLUX.1 [dev] licence). Read the licence of the exact model you run, not the family name.

What the licences ask about labels

Very little. The FLUX.1 [dev] licence asks you to make sure outputs carry a disclosure that AI was used "to the extent required under applicable law" (FLUX.1 [dev] licence), and Black Forest Labs' usage policy forbids content likely to deceive a reasonable viewer without appropriate disclosure, and forbids removing or interfering with C2PA credentials or digital watermarks (BFL usage policy). We found no general output-labelling clause in the Stability AI Community License (Stability AI licence).

In practice, the obligation sits with you and with the laws where you publish. The EU AI Act and platform rules apply the same way to an image made at home as to one made in an app; see real people in prompts.

Watermarks and what your front end saves

Watermarking exists in the reference code but is easy to lose. The original Stable Diffusion script embedded an invisible "StableDiffusionV1" watermark using the invisible-watermark library (CompVis reference script). Hugging Face Diffusers uses that library by default for SDXL only if it is installed, and it can be turned off (Diffusers SDXL docs). Black Forest Labs added an option to embed invisible watermarks in its FLUX.2 code and recommends C2PA metadata (FLUX.2 repository), and says its hosted FLUX.2 Pro API applies signed C2PA metadata (FLUX.2 [dev] model card). That signature comes from the hosted service, not from a model running on your computer.

What local front ends do save is the recipe. AUTOMATIC1111's web UI adds generation parameters to PNGs as a text chunk (AUTOMATIC1111 wiki), and ComfyUI lets you recover complete workflows and seeds from the media it saves (ComfyUI). That is the opposite of the commercial files we tested: no signature, but your full prompt in plain text for anyone who opens the file. We have not yet tested a local PNG ourselves; that test goes on the Stable Diffusion and Flux page.

Adding provenance yourself

If you want a local image to carry a record, add one. Research such as Meta's Stable Signature shows a model's decoder can be fine-tuned so every image it makes hides a detectable watermark (Fernandez et al., 2023), but that is a research method, not a switch in your front end. For most people the practical steps are simpler: keep the original file, keep a prompt log with a file hash, and disclose AI use wherever you publish.

To understand what the model itself does and why it leaves no trace on its own, see how image models work; for what commercial files carry, see what your AI image carries.

A missing label does not mean an image is human-made.

Sources

  1. Hugging Face: Stable Diffusion 3.5 Large model card (accessed 2026-11-02)huggingface.co
  2. Stability AI: Community License Agreement (accessed 2026-11-02)stability.ai
  3. Black Forest Labs: FLUX.1 [dev] licence (accessed 2026-11-02)huggingface.co
  4. Hugging Face: FLUX.1 [schnell] model card (accessed 2026-11-02)huggingface.co
  5. Hugging Face: FLUX.2 [dev] model card (accessed 2026-11-02)huggingface.co
  6. GitHub: black-forest-labs/flux2 (accessed 2026-11-02)github.com
  7. Black Forest Labs: Usage policy (accessed 2026-11-02)bfl.ai
  8. GitHub: CompVis stable-diffusion txt2img.py (accessed 2026-11-02)github.com
  9. Hugging Face Diffusers: SDXL pipeline docs (accessed 2026-11-02)huggingface.co
  10. Fernandez et al.: The Stable Signature, arXiv 2303.15435, 2023arxiv.org
  11. AUTOMATIC1111 wiki: Features, PNG info (accessed 2026-11-02)github.com
  12. GitHub: ComfyUI (accessed 2026-11-02)github.com

Questions

Do open-source image models add watermarks?

Some reference code does, such as the original Stable Diffusion script and Diffusers for SDXL when the watermark library is installed, but it can be switched off and many front ends skip it.

Do Stable Diffusion images have content credentials?

Not when you run the model locally: no company signs the file. Hosted APIs are different; Black Forest Labs says its FLUX.2 Pro API applies signed C2PA metadata.

Can I sell images made with Flux?

The FLUX.1 [dev] licence allows commercial use of outputs except where it says otherwise, and FLUX.1 [schnell] is Apache 2.0. Check the licence of the exact model and the rules of the site you sell on.

Does Stable Diffusion save my prompt in the image?

Popular front ends do: AUTOMATIC1111 writes generation parameters into PNG text chunks and ComfyUI saves the workflow and seed. Check before you share an original file.

Related reading