What to upload: the ten photos that decide everything
Almost every disappointing AI photo traces back to the uploads. What to include, what to leave out, and why ordinary phone snaps in varied light beat your five best photos.
Nearly every disappointing result from an AI photo app traces back to the same place: the ten photos you uploaded at the start. The generation step is doing what it was asked. The face model it was given was the problem.
This is the one part of the process you fully control, it takes about five minutes, and it is worth spending them.
The counter-intuitive part: variety beats quality
The instinct is to upload your best photos — the flattering ones, the ones from the wedding, the one a friend with a real camera took. That instinct produces worse results.
A face model is learning what your face is, structurally, across conditions. It needs to see how your features behave in hard light and soft light, from slightly above and slightly below, with and without a smile. Five excellent photos from one afternoon teach it one lighting setup and one angle. Ten ordinary phone snaps from ten different days teach it your face.
If a photo is in focus and your face is clearly visible, it is good enough. It does not need to be a photo you like.
What to include
- Around ten photos. Fewer than about eight and the model has too little to work with; far more adds little once the variety is there.
- Different days. Different days mean different light, different hair, different clothes — all useful signal.
- Different light. Daylight, indoor light, shade, an overcast afternoon. This matters more than any other single factor.
- Mostly head and shoulders, with your face taking up a decent part of the frame.
- A couple of angles. Straight on, and slightly turned. Not a full profile.
- Both expressions. A neutral face and a real smile. A model trained only on one will insist on it.
- Recent photos. If you want to look like yourself now, do not include the haircut from three years ago.
What to leave out
- Sunglasses, and hats that shade the eyes. The eyes carry most of the likeness.
- Heavy filters and beauty modes. These change facial geometry, not just skin. The model learns the altered face and then produces it, and the result reads as someone adjacent to you.
- Group photos. Even cropped, another face at the edge of the frame is noise. Crop tightly or skip it.
- Anything blurry, very dark, or very low resolution. There is nothing to learn from a smear.
- Photos where you are far away. A face forty pixels wide contributes nothing.
- Extreme angles. Full profile and steep from-below shots pull the model in directions you probably do not want.
- Masks, and hair completely covering the face. For the obvious reason.
Facial hair, glasses and hair length
These are consistency questions rather than rules. A model trained on a mix of bearded and clean-shaven photos will produce a mix, and it will not always be the one you wanted. If you want a specific look, weight the uploads toward it.
Glasses are the same, with one addition: reflections in the lenses in your uploads tend to persist, so prefer photos where your eyes are clearly visible through them. If you sometimes wear them and sometimes do not, include both and expect both back.
Once you have generated something
Judge the set, not the photo. If one result is off, that is normal variance. If every result has the same problem — consistently too young, consistently a different jaw, consistently smiling when you did not want it to — that is the uploads talking, and the fix is upstream. Re-do the face model with more varied photos rather than generating another fifty.
And view candidates at the size they will be used. A portrait that looks superb full-screen can fall apart as a 40-pixel circle beside your name, which is a different job with different rules.
When the face model is right, everything downstream gets easier. Browse the template catalogue to see what is available once it is.
