Which AI keeps your face? 8 image models tested on headshots

We also asked Gemini 3.1 Pro to judge. Rating each headshot alone, it gave nearly all of them “unmistakably the same person,” including ones where glasses had been added, so the scores were useless. Shown all eight side by side, it ranked Nano Banana Pro highest for likeness and put FLUX 3 4th.

The face-recognition score disagreed, and so did a blind comparison in which a reviewer picked between FLUX 3 and Nano Banana 2 without knowing which was which: FLUX 3 looked more like the person in 9 of 12 pairs. Reading the judge’s notes, it rewarded smooth skin and clean light, which is also exactly what makes a face look less like its owner. If you use an AI to grade AI headshots, check it against a face-recognition score.

We rewrote the prompt five times for the top models: removing selfie leftovers, adding studio lighting, insisting on real skin texture. Each change fixed what it targeted, but for FLUX 3 and Nano Banana 2 none moved average likeness by more than about 0.01, close to the variation between two renders of the same prompt. The gap between models was up to 0.17. Picking the model matters more than wording the prompt. Rendering at 2K instead of 1K didn’t help either: it made FLUX 3 less faithful and twice as expensive.

Which AI image model keeps your face best in a headshot?

In our October 19, 2026 test of 8 models on 12 selfies, FLUX 3 kept faces closest to the original (mean likeness 0.89 on a 0–1 face-recognition score), ahead of Nano Banana 2 (0.82). It also cost less than either Nano Banana model, at about 5¢ per image.

Is Nano Banana good for headshots?

Nano Banana 2 and Nano Banana Pro made the most polished headshots in our test and an AI judge preferred them, but both changed faces more than FLUX 3: mean likeness 0.82 for Nano Banana 2 versus 0.89. The usual change was smoothing: wrinkles, moles and skin texture softened.

Do AI headshot generators change your skin colour?

Every model shifted skin colour somewhat. Measured on the cheeks and ignoring brightness, the average shift ranged from 4.1 to 8.3 CIELab units by model, and up to 18.6 on a single photo.

Can an AI judge tell whether a headshot still looks like you?

Not reliably. Asked to rate likeness on a 1–5 scale, Gemini 3.1 Pro gave nearly every headshot 5, including ones where glasses had been added. Ranking side by side worked better, but it still preferred the more polished images, which disagreed with both the face-recognition score and a blind human comparison.

• People. 12 fictional adults aged 25–72, of varied skin tone and ethnicity, with glasses, a hijab, braids, beards and freckles. Each selfie was made to look like a real customer photo: webcams, car selfies, backlight, an arm in frame. Half were generated with FLUX 3 and half with Nano Banana Pro, so neither family only edited faces from its own style; it made no difference to the ranking.
• Task. Each model got the selfie and the same prompt: a professional headshot on a grey studio backdrop, same clothes, face unchanged. One render per model per person, at about 1 megapixel.
• Likeness. Faces found and aligned with YuNet, compared with SFace (both OpenCV Zoo, open models), both images scaled to the same size first. Cosine similarity, 0–1.
• Skin shift. CIELab colour of two cheek patches placed from facial landmarks; we report the a*b* distance and leave out lightness.
• Limits. Twelve people is enough to separate these models, not to prove any model is fair across all faces. Face-recognition scores are known to be less accurate for some groups. Cost is what each render cost us through the provider’s API at 1K, or the list price where the API doesn’t report it. OpenAI’s image models weren’t tested.
• We run DIY Headshots, which makes $1 headshots, and we chose our model with this test. See the methodology.

ai-headshot-likeness-test.json