Finding a resemblance is one decision.
Knowing what to keep takes a few more.
Follow one image through comparison, a proposed identity, visual review and selection for a set.
Narrated teaching example. The portraits, groups and candidate order are authored illustrations. No models or live collection are connected.
The guided example is separate from your own exploration.
Start with the collection, or explore any view. Decisions stay with the image as you move.
The interface can look similar while the decision means something different.
A selected set can support a draft or become source material for further production, through separate choices.
In the inspected generated-image path, a new result retains its intended character and parent-image link. That intention is separate from its assigned identity. New results return to the collection to be assessed.
Changing a reference or rejecting an image does not automatically retrain a model, regenerate an image or publish a post.
Read the Studio piece: Characters that stay themselves.Four written accounts contribute different parts of this explanation.
This is a browser-only simulation based on source inspected on 24 September 2026. It demonstrates the ordinary-reference assignment and stored-candidate rejection path, plus an explicit set-builder eligibility rule. Other assignment routes, manual sets and downstream consumers have different checks.
The drawings do not demonstrate recognition accuracy. Descriptions, grouping and candidate ranking are authored. Actions change this example only, and a reload resets it. No real images, measured vectors, live write acknowledgements or production guarantees are represented. Ara narration uses real audio-section boundaries; word alignment is pending. Real collection media is not included.