Why Some AI Portraits Look Fake (and How to Spot the Good Ones)
A founder-level explainer on what separates a great AI portrait from one that lands in the uncanny valley. The five tells of a weak portrait, the five hallmarks of a strong one, and what to look for at preview time.

Builds AI art tools that real customers actually love. Obsessed with the craft of digital portraiture and the small details that make a portrait feel like a gift.

Every week I see screenshots from customers comparing FrameArto previews to competitor previews, and the question is almost always the same: why does that one look right and that one look wrong? The answer is not magic, it is craft. AI portraits live or die on a small number of details, and once you know what they are you can spot a strong one in under five seconds. This is the founder primer on quality, written so that you can judge any AI portrait you see, ours or anyone elseβs.
The five tells of a weak AI portrait
1. The eyes are off
The single most reliable tell of a weak AI portrait is the eyes. Either they do not match each other (different size, different focus, different gaze direction), or they are vacant in a way real eyes never are. Real eyes have a small bright catchlight, a clear iris boundary, and a sense of the person being home behind them. Weak AI flattens all three. Look at the eyes first. If they fail, nothing else matters.
2. Hands and paws are wrong
AI has historically struggled with hands. Extra fingers, fused fingers, weirdly bent thumbs. Pets get the same treatment with paws and toes. A strong portrait will either render hands and paws accurately, or it will compose them out of the frame. If a portrait shows hands and the hands look subtly off, the rest of the portrait is probably subtly off too. The good ones do not get away with hand mistakes.
3. The face is symmetric in a way real faces are not
Real faces are subtly asymmetric. One eye sits a hair higher, one corner of the mouth pulls a hair further, the nose has a slight lean. Weak AI flattens this asymmetry into a doll-like mirror. The portrait reads as off because the source person was asymmetric and the rendering is not. Look for tiny asymmetries. Their presence is a sign the AI honoured the actual face.
4. The skin texture is wrong for the style
Watercolor skin should look like watercolor: visible paper texture, soft pigment edges, gentle paint pooling. Oil skin should look like oil: thick paint, visible brush direction, slight glossiness in the highlights. If the skin in either style looks like the AI applied a filter to a photograph, the rendering is shallow. A strong portrait makes you forget there was ever a photograph.
5. The background does not match the subjectβs lighting
A common AI failure is the subjectβs light direction not matching the background. The face is lit from the left, the wall behind is lit from the right. This is almost invisible on a phone screen and obvious on a wall. Look at the shadows on the face versus the shadows in the background. They should agree. If they argue, the portrait will feel uncanny once printed.
The five hallmarks of a strong AI portrait
- Eyes that have a catchlight, a clear iris, and a sense of presence.
- Hands or paws that are anatomically right, or composed out of frame on purpose.
- Subtle facial asymmetry preserved from the source photo.
- Skin and coat texture that matches the medium of the style (paper, paint, ink).
- Lighting that agrees between subject and background, with consistent shadow direction.
βA useful at-a-glance test. Cover the bottom half of the portrait with your thumb on the phone screen. Look only at the eyes. If you can tell the portrait is of the right person from the eyes alone, it is a strong portrait. If you cannot, no amount of detail elsewhere will save it.β
How FrameArto handles the hard cases
We invest specifically in the failure modes above. Our prompt structure includes constraints around eye accuracy and asymmetry preservation. We re-run hands and paws through a refinement pass when they appear in frame. And every preview goes through an automated check before it shows up in the gallery. None of this makes us perfect, but it pulls us out of the uncanny valley most of the time, and on the rare occasion a preview lands wrong, the free preview lets you regenerate without paying.
What to do at preview time
Spend thirty seconds looking, in this order: eyes, then hands or paws, then asymmetry, then skin texture, then background lighting. If all five hold up, print. If two or more fail, regenerate. If only one fails, decide whether it matters for the recipient. A slightly off paw on a watercolor pet portrait is forgivable. An off eye on a memorial portrait is not. Match the tolerance to the occasion.
The honest bar
AI portraits at our level are not perfect. They are, however, often better than the average human commissioned portrait at twenty times the price. The bar I want every customer to apply is simple: does this portrait look like the person or animal I love? If yes, print it. If no, regenerate. We will keep working on closing the remaining gap. Most of the perceived weirdness in AI portraits will continue to disappear in the next two model generations.
Frequently Asked Questions
The questions readers ask us most about this topic.
Why do AI portraits sometimes look "off" but I cannot say why?
Usually one or two of the five tells (eyes, hands, asymmetry, texture, lighting) is failing subtly. The brain registers the failure faster than the conscious mind names it. Look at the eyes first, that is where most of the perceived weirdness lives.
Are some AI services better at hands than others?
Yes, by a significant margin. Hands and paws are where the gap between competent and great AI portrait services is widest. We refine hands and paws through a dedicated pass before previews surface. Generic AI image services typically do not.
Can I ask the AI to redo just the eyes if everything else looks good?
Yes. The custom portrait flow accepts targeted feedback. "Keep everything else, redo only the eyes to look more like the source photo" is a normal brief and we honour it without recharging.
Will AI portraits eventually be indistinguishable from photographs?
They already are for casual viewing in many styles. Forensic inspection still reveals them. The interesting question is not whether they will become indistinguishable but whether anyone will care. Most customers do not. The portrait either looks like the loved one or it does not.
If I notice something off in my FrameArto preview, what should I do?
Regenerate. The free preview is free, and the second attempt almost always cleans up the failure mode the first one had. If two regenerations both feel off, switch styles. The photo may be fighting the style rather than the AI fighting the photo.
See what your photo can become.
Generate a free preview in three styles. No credit card, no commitment, results in under three minutes.
Start My Free Preview

