Every product in this category prices per run or per image. Both are the wrong unit, and the right one is easy to compute and almost never published.
The metric
cost_per_usable = total_cost / outputs_you_would_actually_ship
That denominator is the keeper rate, and it varies far more between products and between faces than price does.
A run returning 100 images at a 4% keeper rate gives you four usable photographs. A run returning 30 at 40% gives you twelve. If the first is half the price it is still worse per unit, and it also costs you the time spent looking through 96 rejects.
Nobody publishes keeper rate, including us. It is the number I would ask any vendor for, and the answer tells you something either way.
Why the denominator moves so much
Two things dominate it, and neither is the product.
Input variance. Fifteen photos from one afternoon collapses the keeper rate on every product, because the model learns the session rather than the person.
Which face. Keeper rate on an easy face is high everywhere. On a hard one it separates products by a factor of several. Averaging across a mixed team hides that, which matters when the low scores cluster on the same people every time.
Measuring it without instrumenting anything
Run one subject through two products the same afternoon, then have someone who is not the subject mark each output ship or no-ship. The subject is a bad judge of their own likeness and a worse judge of whether a photograph is usable.
Two numbers, five minutes, and it reorders most published rankings.
The comparisons, read for method
We write these about competitors so treat the verdicts as claims. What transfers is the criteria.
Against HeadshotPro and against TryItOnAI both set out the inputs used, which is the part worth checking in anyone's comparison.
On the pricing tiers specifically: free versus paid is mostly about output volume rather than resolution, which is exactly the keeper-rate problem. And doing this cheaply without wasting the money covers where to economise, which is not on the number of outputs.
The case with a hard constraint
Worth flagging because it is unusual in this category: a CV photograph is governed by local convention rather than image quality, and the convention inverts across borders. Mandatory in much of continental Europe, a liability in the UK and US. Current practice.
That is a case where the best possible output can still be the wrong decision, which no metric catches.
What I would log if I were integrating this
outputs_returned, outputs_downloaded, time_to_first_download. Three fields, and between them they give you keeper rate and the latency users actually feel. Neither appears on any vendor's pricing page, ours included.
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
For two tools priced within a few dollars of each other, the BetterPic and HeadshotPro comparison.
Disclosure: I work on one of the tools priced in that comparison, BetterPic.
Top comments (0)