AI disclosure

How the images are made

Everything published here is generated with ComfyUI, running locally on my own hardware, from text prompts I wrote. Nothing is a photograph. Nothing is a retouched photograph. No photograph of a person is used as an input, a reference or a starting frame.

If you are signed in, the placard beside each picture prints the actual settings that produced it — the model, the prompts, the sampler, the seed, the LoRAs. It is read out of the file the generator wrote, not typed in afterwards, so it is a record rather than a claim.

The people in them

Every person depicted on this site is fictional and computer-synthesised. They have names, biographies and portfolios because that is the conceit of the project. None of them is a person. None of them has ever been photographed, because there has never been anything to photograph.

No image here depicts a real, identifiable person. I do not use models, LoRAs or embeddings trained on a named individual, and I do not prompt for anyone by name. A generative model is built from a very large number of faces and can produce one that resembles someone; if a face here looks like a person you know, that resemblance is coincidental and unintended.

If you believe an image resembles a specific real person closely enough to be a problem, tell me and I will take it down. I would rather lose a picture than argue about a face.

Report an image

Transparency

Article 50 of the EU AI Act requires that artificially generated image content be disclosed as such, clearly and distinguishably. Whether or not this site falls within its scope, the intent is easy to honour and I would rather over-disclose than argue about jurisdiction:

• Every page on the site carries the disclosure in its footer.
• Every image page repeats it beside the picture.
• The name of the site says it, the about page says it, and no image anywhere is ever presented as a record of something that happened.
• The generation settings for each image are published rather than hidden.

Machine-readable provenance marking — C2PA content credentials in the files themselves — is not implemented yet. The published derivatives are deliberately stripped of all metadata, because the original files carry prompts and settings that should not travel with a downloaded picture, and signed credentials would have to be added back deliberately rather than left in by accident. It is on the list.