Everyone tries this in ChatGPT first, and here is where it falls over

The first attempt most people make now is in a general chat assistant. It is free, it is already open, and the output looks impressive at first glance.

It is also a structurally different tool, and knowing why explains the disappointment that usually follows.

The architectural difference

A dedicated generator trains a small adapter on fifteen to twenty photographs so that a trigger token means your face. There is an explicit identity-learning step, which takes real compute and real time.

A chat assistant conditions on one or two reference images at inference. No training step, no learned identity. Fast, free, and correspondingly weaker at likeness.

That is the whole explanation for "it looks like my cousin". It is not a quality gap, it is a missing stage. Where it falls over in practice goes through the specific failure modes.

The second consequence people hit later

Consistency across a set. Without a learned identity, each generation drifts, so six outputs meant to look like one session do not.

For a single stylised avatar that does not matter. For a professional set, or forty people on a directory, it is the whole requirement.

Where the general tools are genuinely fine

A one-off illustration. Something playful. Anything where recognisability is not the point. There is no reason to pay for those.

And where they are not

Anywhere a stranger has to recognise you from the result: a profile in a recruiter's list, a clinician's page, a company directory.

The test is the same regardless of which route you take. One hard face, ideally your own. Fifteen photographs across four different days. Then ask someone who knows that person whether it still looks like them.

Two useful pieces of context

The corporate side of this arrived faster than anyone expected, which is why the tooling now splits between individual and organisational use: what generated portraits are covers the general case.

And we acquired AI Model Agency last year, which is where a good part of the volume experience behind these observations comes from. Worth knowing when reading anything we publish about competitors.

The honest summary

Start in a chat assistant if you want to see the shape of the thing. Do not conclude the category is bad from it, and do not put the result on a profile where someone needs to recognise you.