Producing one good portrait and producing forty that belong together are different problems, and the second is not the first repeated.
Why the metric changes
A grid where every image is individually good and collectively mismatched reads worse than a grid of merely adequate ones that match. The eye is quick at spotting a set and reads inconsistency as carelessness about everything else.
So above roughly ten subjects you are optimising variance, not mean quality. Four dimensions drift, and all four are invisible one image at a time:
- background tone
- crop distance
- colour temperature
- eye-line height in frame
Measuring it cheaply
Put the outputs in a grid at 96 pixels and look at the grid rather than the images. That is the rendering condition of a real directory and it makes all four obvious in about five seconds.
If you want a number rather than an impression, the cheap proxies are: mean pixel value of a fixed background patch, detected face-box height as a fraction of frame, and mean colour temperature of a skin-region sample. Standard deviation across the set on each. You do not need them to be accurate, only comparable within one batch.
The failure that matters most
Check whose outputs you would not use. If the rejects cluster on the colleagues with the most textured hair or the deepest skin tones, you have found something that does not improve with scale, and averaging across the team hides it.
That is the single most important check in a team rollout and it is almost never run.
The eight-person pilot
Pick eight people who look as different from each other as your organisation allows. Same day, same settings. Grid at 96 pixels. Check the rejects against the people.
An afternoon, and it predicts most of what goes wrong at 200.
The part that is not technical
The reason company photography decays is scheduling. Forty calendars, one room, six people who reschedule twice, and anyone remote making it impossible rather than merely hard.
That is why a session solves the problem once and the grid degrades one new joiner at a time. The variable that decides the approach is turnover: under a tenth a year, book a photographer; over it, you need something reproducible for people who have not arrived yet.
What photographers charge is the baseline, and the team-scale write-up plus the version aimed at the operational side cover the alternative.
Remote teams are the sharpest case, where there is no shared room or light at all: employee photographs for distributed teams and the virtual-session version.
The one thing I would instrument
Store the source and every derived crop separately, with the settings used. When the design changes and you have to regenerate 200 variants, having the originals is a job and not having them is a re-shoot.
Disclosure: I work on one of the tools measured that way — BetterPic.
Top comments (0)