Search for a portrait generator and you get four architecturally different products under one label. Picking the wrong category is more costly than picking the wrong product within a category.
1. Fine-tuning generators
Upload fifteen to twenty photographs, a small adapter is trained so a trigger token means your face, then it runs against a prompt library.
Best likeness, because there is an explicit identity-learning step. Slowest, because training takes real compute. Highest input sensitivity: fifteen images from one afternoon teaches the model the shirt.
This is the category most people mean, and the LinkedIn-focused ones are the narrow version of it.
2. Reference-conditioned, no training
One or two images conditioning a generation at inference. Fast and cheap, no training step, and correspondingly weaker likeness. Fine for a stylised avatar, unreliable for something meant to be recognisably you.
Most in-chat image tools sit here, which explains the common experience of "it looks like my cousin".
3. Editors
No generation at all. Crop, resize, background replacement, colour correction. Underrated, because a large share of "my photo doesn't work" turns out to be a crop problem with an existing photograph rather than a need for a new one. The profile picture tools mostly live here.
4. Business image sets
A different brief: people plus workspace plus product plus detail. Confusing this with a headshot is the most expensive common mistake in the category, because you book one and need the other. Business photo generation and the roundup of those tools cover it.
How to tell which you are being sold
One question: does it train on your uploads, and how many does it ask for?
Fifteen to twenty means category one. One or two means category two. Zero means category three or four. The number of requested inputs tells you the architecture, and the architecture tells you what likeness to expect.
Which one you need
If a stranger has to recognise you from the result, category one, and test it on a hard face first. A free entry point is the cheapest way to find out whether your face is one of the difficult ones.
If you have a decent photograph that renders badly, category three, and you can probably fix it in five minutes.
If you need a website's worth of imagery, category four, and a headshot tool will disappoint you no matter how good it is at headshots.
Related, on our comparison blog: Free AI headshot generators: what free actually means
A longer comparison, with eleven tools side by side: The best AI headshot generator in 2026: four tests that actually decide it
Disclosure: the category-three tool I work on is BetterPic's AI headshot generator.
Top comments (0)