Six roundups, six winners: what the disagreement is actually about

Read four comparisons of portrait generators and you get four different winners. It looks like corruption, and part of it is: this category runs on affiliate commission to an unusual degree.

But a large share of the disagreement survives even when everyone is honest, and it comes from four decisions each reviewer makes without stating them.

Whose faces

The biggest variable by a distance. These models fail unevenly across hair textures, skin tones and features, and that tail is where products separate.

A review run on three cooperative faces has sampled only the easy part of the distribution. It will find a tie and report a winner.

What counts as good

Photographic quality and likeness are separate axes, and products trade one against the other. Heavier stylisation raises the first and lowers the second.

A reviewer weighting "is this a nice photograph" ranks differently from one weighting "does this still look like the person". Both positions are defensible; neither is usually declared.

How the inputs were prepared

Fifteen images from one session degrades every product, because the model learns the session rather than the subject. A reviewer who does that is measuring graceful degradation, which is a real property and not the claimed one.

When it ran

Queue depth varies with load. A turnaround figure from a quiet Sunday and one from a Monday morning describe the same product differently.

Reading ours with that lens

We write these about competitors, so treat the verdicts as claims and the criteria as the transferable part.

The dimension-level pages exist because products genuinely rank differently by brief: executives against Aragon is a narrow case with a longer approval loop and more scrutiny, which some tools handle and others force you to work around.

The overview pages are against Aragon, against HeadshotPro and against InstaHeadshots. And our roundup of the field is the broadest, and the one where our conflict of interest is largest.

The test that outranks all of it

One hard face, ideally your own. Fifteen photographs across four different days. Two products on the same afternoon. Then mix five generated images with five real ones and ask someone who knows that face to separate them.

Twenty minutes, no commission involved, and it samples the only part of the distribution you actually care about.

What I would want the category to publish

Keeper rate: usable outputs over total outputs, per product, on a stated test set. It is the number that decides whether people are satisfied, and nobody publishes it, us included. That is a fair thing to hold against all of us.