Written by the BetterPic team. Disclosure: BetterPic sells one of the tools mentioned here. This piece is about the input side, which no vendor writes about because there is nothing to sell in it.
Most comparisons of AI headshot tools argue about the model. In practice, roughly half the variance in what comes back is decided before the model runs at all, by the twenty photographs you hand it. Two people using the same tool on the same day get results a grade apart because one of them uploaded a usable set and the other uploaded a camera roll.
Variety in angle, not in situation. The model needs to learn your face in three dimensions, so it wants you turned slightly left, slightly right, straight on, chin a little up, chin a little down. It does not need you at a wedding, at a beach and in a car. Different rooms add noise; different angles add information.
One face per photo, and it has to be yours. Group shots with you cropped out are the single most common contaminant. If another face survives anywhere in the frame, some of it ends up in you.
Recent, and consistent with each other. A set spanning four years teaches the model two different people and it will average them. If you have changed your hair, your weight or your glasses, discard everything from before.
Neutral and varied expressions. Include some with a closed mouth. Teeth are the most common reason people reject an otherwise good output, and a set that is all broad smiles gives the model nothing else to work from.
Heavy filters, including the invisible ones. Phone beauty modes are on by default on a lot of handsets and they smooth skin texture before the file is even saved. The model learns the smoothed version, and the output looks like a relative of yours.
Sunglasses, hats, and hair covering the jaw. Anything that occludes the face boundary teaches the model an outline you do not have.
All the same lighting. Twenty photos taken in the same room under the same lamp give the model one lighting condition, and it will bake that condition into everything it produces.
Very low resolution or heavy compression. Screenshots of photos, images pulled from messaging apps, anything that has been re-saved several times. The artefacts are learned as if they were features of your face.
Put your candidate set in one folder and look at the thumbnails together. Three questions:
If the answer to the last one is no, add three-quarter angles until it is yes. That single fix improves more results than switching tools does.
Every vendor publishes turnaround time and image count because those are easy to measure and easy to improve. None of them publish yield — how many of the batch you would actually use — because it depends on your face and on what you uploaded, and it is the only number that decides whether the purchase was worth it.