How to Take Selfies for an AI Headshot Generator

Disclosure: XPortrait is our AI headshot product. We will tell you when a competitor tool or a professional photographer produces better results for your situation.

Most failed AI headshots are not the tool's fault. The model is only as good as the photos it trains on. Upload ten near-identical bathroom-mirror selfies under flat overhead light and you will get a batch with poor likeness, muddy skin tones, and artifacts around your hair — then spend time blaming the generator. Fifteen clear, varied, well-lit photos reliably outperform fifty blurry near-duplicates, across every major tool.

This guide covers exactly what to shoot before you pay for a generation. The whole setup takes about 20 minutes.

What the AI is actually learning from your photos

When you upload photos, the tool runs a fine-tuning process: it teaches a generative model what your face specifically looks like — bone structure, eye shape, skin texture, hair pattern — by learning from the variation across your uploads. It is not averaging your selfies or pasting your face onto a template. It is constructing a compact representation of your appearance that it can then apply across different lighting setups, outfits, and angles.

This explains why variety matters more than quantity. If every photo is the same front-facing close-up with the same flat overhead light, the model learns one narrow slice of your face — and the output drifts when it tries to render you from any other angle or under different lighting. Identity drift — where the output looks like a plausible stranger rather than you — is almost always a training-photo problem, not a tool failure.

How many photos to upload

Most tools ask for 8–20 photos. XPortrait works well from 8–12; tools that do more extensive fine-tuning — BetterPic, Aragon — benefit from 15–20. The floor is 8: below that, the model has too little variation to learn from and likeness failure rates rise sharply. Above 25, you get diminishing returns; the variety gains plateau and you are just adding upload time.

The angles to cover

Cover these four angle categories across your upload set. You do not need equal numbers in each — two or three from each is enough.

1. Front-facing, straight-on. Your face centered, looking directly into the camera. Three or four of these.
2. 3/4 angle left. Turn 30–45 degrees to your left. This gives the model your nose bridge, the depth of one ear, and the shadow side of your face.
3. 3/4 angle right. Mirror of the above.
4. Slight chin-down or chin-up. One looking slightly down toward the lens, one slightly up. This helps the model learn your facial proportions across viewing angles rather than a single flat plane.

Skip full profile (90 degrees) — it is too extreme and adds noise rather than useful data. Looking over your shoulder produces warped face proportions the model cannot normalise cleanly.

Lighting: the variable with the highest impact

Lighting in your training photos directly determines the skin-texture detail the model has to learn. The model absorbs shadow gradients from your photos — if your face is flat-lit with no shadow variation, the generated outputs render your face flat too. If your face is backlit and dark, the model infers the wrong color values for your skin.

• Soft window light is the best option. Sit facing a window during daylight hours — not in direct sun, but in the indirect light coming through it. Morning or late afternoon gives a gentle directional cast that teaches the model your face's three-dimensional structure.
• Include photos in 2–3 different lighting conditions across the batch. One slightly warmer (golden hour), one neutral (overcast or open shade). Variety here helps the model generalise rather than locking in a single lighting assumption.
• Avoid harsh overhead lights. A ceiling bulb directly above you creates deep shadow under your eyes and nose — features the model reproduces in outputs even when the generated lighting is correct.
• Avoid backlighting. Standing with a bright window behind you underexposes your face. The model reads the darkened version as your actual skin tone.
• Avoid phone front-camera flash. It produces reflected, flat light that kills skin texture and overexposes the center of the face.

Background: plain is enough

The model needs to isolate your face from the background to learn your features accurately. A cluttered background — bookshelf, textured wall, other people in the frame — forces the model to do more separation work and leaves room for error at your face edge.

A plain white or light grey wall, a solid-color door, or a large blank surface works. The background will not appear in your headshot output at all — the AI generates a new professional background during generation. You only care about your training photo background to the extent that it helps the AI see your face cleanly.

One practical rule: do not shoot against a background that is close to your skin tone. A pale wall against a fair complexion, or a very dark wall against a deep complexion, reduces the contrast the model needs to define your face boundary cleanly.

What to wear (and why it barely matters)

Most AI headshot tools generate new clothing in the final output — the model learns your face, not your shirt. What you wear in training photos has minimal effect on the output wardrobe, which is selected separately during generation. Two rules still apply:

• Avoid pure white. It reflects light and saturates in overexposed smartphone photos, bleaching out fine skin texture.
• Avoid pure black. It absorbs light and reduces the shadow variation the model needs to render realistic skin depth.

A plain grey, navy, or any medium-tone solid is fine. More important than color: do not change your hair, grow or shave a beard, or wear glasses in some photos and not others. Appearance changes across uploads confuse the model and cause likeness drift.

Expression: neutral, not posed

Your expression in training photos does not become the expression in outputs — the tool generates a new expression based on the style you select. But exaggerated expressions in training photos do cause problems. A wide-mouthed laugh creates unusual face proportions — cheeks puffed, eyes narrowed — that the model treats as a fixed characteristic rather than a momentary state. Outputs can inherit slightly puffed cheeks or narrower eye apertures than you actually have.

Use a neutral, closed-mouth expression or a small natural smile across most training photos. The right frame is "hold still for a second" — calm, not posed.

Nine things that will ruin your batch

• Instagram, Snapchat, or any beauty filters — they alter your face geometry. The model learns the filtered version, not you.
• Sunglasses or any eye covering — eyes are the highest-weight feature in portrait rendering. Any photo with covered eyes should be excluded.
• Hats that cut off your hairline — the model will either hallucinate your hair or produce outputs where the hairline looks wrong.
• Group photos with other people in frame — the model tries to learn multiple faces and produces composite artifacts.
• Blurry or out-of-focus photos — even two blurry uploads in a batch of twelve degrade sharpness across the whole batch.
• Old photos that no longer represent your appearance — if you changed your hair, gained or lost significant weight, or aged noticeably since the photos were taken, the model produces outputs of a past version of you.
• Extreme camera angles — looking straight up into a tilted phone or sharply down at 60 degrees gives the model face proportions it cannot normalise.
• Cropped faces — if the top of your head or your chin is cut off, some models struggle to extrapolate the missing area and produce artifacts at the edge.
• Brightness-boosted dim photos — brightening a dark photo in editing creates noise and grain that the model reads as skin texture.

When the AI still fails despite good photos

Good training photos do not guarantee good output. Three failure modes persist even with excellent inputs:

• Very curly or highly textured hair — every major AI headshot tool still produces inconsistent results here. The model tends to smooth or regularise texture, producing hair that looks altered. This is a model limitation, not a training-photo problem you can solve by shooting differently. BetterPic handles it slightly better than the category average, but no tool has resolved this consistently.
• Prescription glasses with tinted or reflective lenses — the model cannot cleanly separate the eye from the lens reflection and produces distorted eyes. Generate one batch without glasses as a backup to compare.
• Identity drift on faces under-represented in Western AI training data — where the output gradually looks less like you. Use a free preview to check likeness before paying for a full batch. XPortrait offers a free preview before payment; HeadshotPro and Aragon do not.

TL;DR — the 20-minute shooting checklist

1. Sit facing a window. Daylight, not direct sun.
2. Shoot against a plain wall in a color that contrasts with your skin.
3. Take 12–15 photos: front-facing, 3/4 left, 3/4 right, chin-down, chin-up.
4. Include 2–3 lighting setups across the batch if possible.
5. Wear a medium-tone solid. Not pure white, not pure black.
6. Neutral expression or small natural smile.
7. No filters, no sunglasses, no hats, no other people in frame.
8. Use photos from the last 6 months only.
9. Delete blurry, cropped, or heavily edited photos before uploading.

The model cannot recover what was not in the photos to begin with. Twenty minutes of preparation on the shoot side consistently produces better output than spending the same time reviewing a bad batch.

Founders and solo operators who manage their own photo workflow — without a photography coordinator or comms team to handle logistics — benefit most from getting the input photos right the first time, since a single well-shot batch can serve LinkedIn, investor bios, and conference profiles simultaneously. The output requirements specific to that context are at /en/ai-headshots-for-founders.