DEV Community

Miriam Alonso
Miriam Alonso

Posted on Edited on

The reader is running a classifier and you are choosing the features

There is a useful way to think about a professional photograph that is not about aesthetics: the viewer is running a very fast, very cheap classification, and everything in the frame is a feature.

That framing makes the decisions concrete, because you can ask what each feature contributes rather than whether the image looks nice.

The features that carry weight

Hair-to-background contrast determines whether the head has an edge at small sizes. Get this wrong and the classifier has nothing to work with.

Camera height encodes dominance. Below eye level reads as authority, above as approachability, and the gradient is steeper than people expect.

Hand visibility encodes openness. Visible and relaxed reads open, folded reads defended, hidden reads formal.

Colour temperature encodes warmth almost directly, which is why the same face looks clinical at 5600K and friendly at 4000K.

None of those is subjective. All four are measurable and all four are decided before anyone looks at the image.

The label being predicted differs by context

This is where generic advice fails. A client choosing a lawyer is classifying for risk. A patient choosing a clinician has already accepted competence and is classifying for approachability, which inverts the weight on camera height and hands.

That is not a stylistic preference, it is a different target variable. What a clinician's photograph has to do is close to the inverse of the standard advice, and applying the standard advice there costs bookings.

There is real work on how quickly and on what basis people form these judgments: the psychology of profile images covers what is actually established versus what gets repeated.

Background as a deliberate choice

The most under-considered feature, and the one with the clearest rule: contrast against the subject's hair, not a colour you like. Choosing a background properly goes through it by hair colour, which is the useful axis.

Worked examples beat descriptions

For something this visual, six examples with what each is doing is more useful than any amount of prose, including this.

Two notes on the source

What generated portraits are covers the mechanism if you are new to the category, and we acquired AI Model Agency last year, which is where some of the volume experience behind these observations comes from.

The check

Reduce the image to 48 pixels and ask which of the four features survives. Whatever is left is what the classifier is actually using, and it is usually less than people assume.

Disclosure: I work on BetterPic, which sells against the query best AI headshot generator.

Top comments (0)