Make an Illustrated Avatar

How it works

Your photograph goes to an image-editing model, at whatever size you uploaded it, with an instruction describing a polished personal avatar portrait: a clean illustrated finish, recognizable facial structure, a flattering but faithful expression, and a simple coherent background, followed by a plain instruction not to change identity.

A second, negative instruction leads with changed identity and distorted face, then extra limbs, malformed hands, unreadable text, and watermarks.

A new picture is generated at exactly 1024 by 1024 pixels. Each run starts from fresh noise, so every attempt is a different avatar.

Three guards and one number

It is worth counting how hard this workflow tries to keep your face.

The positive instruction asks for recognizable facial structure. It asks for the expression to be faithful. It then states, separately and unusually, that identity must not change. The negative instruction opens with changed identity and names distorted face immediately after. No other workflow on this site protects one property in that many places.

And the strength is the same one used by the oil painting page.

That is the honest situation. Instructions are guidance — they weight the generation towards some outcomes and away from others, but nothing enforces them. Strength is not guidance; it decides how far the image is permitted to travel from your photograph, and here it is set high because a low setting would return your photo with a slight sheen rather than an illustration.

So the workflow is asking for a large stylistic change and no substantive one, using a control that does not distinguish between them. The guards make drift less likely. They do not make it impossible, and the more unusual your face is, the more the model's idea of a face pulls against it.

Generate several, then choose

This follows directly, and it is how the page is meant to be used.

Because each run starts from new noise, the variation between attempts is not noise in the trivial sense — it is genuinely different interpretations. One will catch your face and another will produce a plausible stranger with your haircut, from the same input, with nothing changed.

That means a single generation tells you very little. Run it several times and treat the set as candidates.

The selection criterion matters too. The temptation is to choose the most appealing image, and the most appealing image is frequently the least faithful, because the drift that makes it appealing is drift away from the specifics of an actual face. A better test is to show two or three to somebody who knows you and use whichever they name without pausing.

Where the reference does the work

The model knows only what your photograph shows it.

A frontal, evenly lit face gives it the full structure — the width across the cheekbones, the set of the eyes, the line of the jaw. There is little left to infer, so there is little room to infer wrongly.

A three-quarter angle hides half of that. Hard side lighting hides more. Glasses with reflections, a hand near the chin, hair across one eye, a strong expression mid-speech — each of these removes information that then gets filled in from the model's general sense of how faces go.

None of that filling in is visible in the result, because the output is a confident, finished illustration either way. It looks equally certain whether it was reconstructing or observing.

An unglamorous, well lit, straight-on photograph is the correct input, however little you like it.

Publication gate

This page ships once the workflow has been run several times each against a frontal portrait, a three-quarter portrait, a portrait in hard side light, and a subject wearing glasses, with likeness assessed by someone who knows the subject rather than by the operator.