Check an AI-generated portrait before you post it

The image appears in the result frame. At a glance it's fine — better than you expected, even. Your finger is already on its way to the download button and you're thinking about where it'll go. Then something vague holds you back, a sense that something's off that you can't quite point to. That feeling is usually right. The problem is knowing where to look.

There's no "before" picture, so your eyes are the check

When you run a photo you took through an editing tool, the result screen gives you two versions to compare: before and after. Your eyes have an anchor. Anything that changed is something the tool touched, and all you have to decide is whether you can live with it.

Generate photorealistic images is different for a simple reason: there was never an earlier version. It takes words only — there's no place to put a photo in — so the picture you're looking at is the first and only version from that run, and the result screen shows just that one image. Nothing is being hidden. There's simply nothing to set beside it.

That has a consequence people don't expect: the face is not a useful test. In a 2022 study published in PNAS, Sophie Nightingale and Hany Farid asked people to sort faces into real and AI-synthesized. In the authors' words, "The average accuracy is 48.2%" — close to the 50% you'd get by guessing. A second group, given training and feedback as they went, did only a little better: their average accuracy "improved slightly to 59.0%" in the study's words. Participants also rated the synthetic faces as slightly more trustworthy than the real ones. That study used the face generators of its day, so it says nothing exact about any tool now, but the lesson holds up: "the face looks real" proves very little. The Northwestern University researchers behind a field guide to spotting AI images add a related warning: "it is not always possible to readily identify artifacts and implausibilities in images, especially portrait images."

So the final check is your own eyes, and eyes need an order to follow. This one goes by distance: the whole image at the size people will see it, then hands and small objects, then the places where two things touch, then any text. The closer you look, the smaller the mistakes get — and the harder they are to explain away.

If your image came from a tool that takes a photo in and changes part of it, you do have a "before," and a different set of problems to watch for.

Look at it the size people will see it

The first job isn't looking hard. It's looking at the right size.

The result frame shrinks the image to fit the page, so judging it there means judging a preview. Once the image has loaded, there's a button labeled View full size to save it. Open it and you see the picture as large as your screen allows. The gap between those two sizes is bigger than most people expect, and nearly everything worth worrying about lives in that gap.

Full size isn't the finish line, though. What counts is the size other people will see, and that's set by where you post, not by you. As a cover image for a group, most viewers get a thumbnail on a phone; an odd hand or a misfolded sleeve disappears at that size, and fixing it earns you nothing. The same image on a product page is another story. Shoppers tap to zoom, and they zoom in exactly where they're suspicious.

So before you inspect anything, answer one quick question: where will this image live, and at what size will people see it? The answer saves you from fixing what nobody will notice — and, more importantly, from skipping what everybody will.

Hands, fingers and small hardware

Now it's time to zoom in, and the question is where.

Kellogg Insight, Northwestern's research magazine, sums up one finding plainly: "while AI models are generally good at creating realistic-looking faces, they are less adept at hands. Hands may lack fingernails, merge together, or feature odd proportions." The same piece extends that to small objects: "Zoom in on details of objects, such as buttons, watches, or buckles." The research guide behind it gives the reason in one line: "Smaller details in an image allow for more opportunities for error."

So the first places to check are wherever small details crowd together — hands and fingers, glasses, jewelry, a clasp on a bag, buttons, small things sitting on a table.

Keep one caveat in mind. Kellogg's write-up also notes that "An extra finger or a missing limb does not automatically imply an image is fake," because real hands vary too. For your purposes, that's a reminder of what this list is: places worth a look, not places that are always wrong. Plenty of images get through every one of them cleanly. The point is to spend your attention where it's most likely to pay off, instead of scanning the whole frame evenly and getting tired before you find anything.

How you look matters as well. Don't sweep across the image like a page of text. Jump straight to each spot and hold still on it for a few seconds before moving on; it's much easier to catch something odd while your eyes are resting on a spot than while they're sliding past it. And if you find yourself pausing on the same spot twice, there's usually something there. A stranger will snag on it too — they just won't stop to work out why.

Where two things touch

The second group is anywhere two things meet. Kellogg's write-up puts it this way: "When there's interactions between people and objects, there are often things that don't look quite right," and gives an example anyone can check: "Sometimes a backpack strap just looks like it merges into your shirt." The research guide makes the same point by comparing two of its examples: "The portrait image has no obvious artifacts, while the full-body image has obvious errors in the hands and around the collar."

These are the spots where an image has to decide what's in front of what, what covers what and exactly where an outline ends — decisions a person drawing by hand would stop and think about. An image made in one pass gets no second look at those decisions before it reaches you. So check a hand holding a bag handle, a strap running over a shoulder, hair against a collar, a chair leg meeting the floor and the edge of a sleeve against skin.

For clothing, stay at the level of the picture itself: do the folds follow the body, do the seams run unbroken, does the fabric meet skin the way that fabric would? People can't name those things, but they notice them right away. Those details decide much of whether an image feels real.

Read every letter, up close

This one is small, costly and easy to skip.

If there's text anywhere in the frame — a sign behind the person, a label on a package in their hand, lettering on a shirt, a storefront in the distance — read every word of it yourself. Out loud if you can, and slowly.

The reason is what you're actually holding: a flat picture in which the text was drawn, not typed. There's no text layer and no box to edit, only shapes that look like letters. Sometimes they spell what you wanted; sometimes they only look like words from a distance and fall apart into nonsense when you zoom in. The Northwestern guide treats this as its own kind of error: "Incomprehensible text and logos as well as overrepresented features are also a distinct functional implausibility generated by AI."

There's a familiar trap here. Small, slightly soft text always looks right to you, because your eyes fill in what you expect from context — you know what that sign is supposed to say, so that's what you read. A stranger doesn't have that head start. Zoom all the way in before you read, not at the size of the whole frame.

And if the image is going to carry your brand name, a misspelled word is the kind of mistake people remember and repeat. It's far harder to forgive than a crooked fold.

This is also where looking runs out. Broken text doesn't get fixed by looking again. You have two options: change the prompt and run it again, or compose the shot so the part with text isn't in the frame.

When something's off: change the prompt, then run it again

The last question is always "now what?" — and the honest answer is on the input side, not in the image you already have.

There's no slider to bend a finished image into shape; it's fixed. What you can change is the prompt that goes into the next run. On the result screen, Generate again keeps your prompt and your level for another variation. Write a new prompt clears the prompt box so you can start over, so if you only want to adjust what you wrote, copy it before you press that one.

There's a smarter way to adjust it, too. Change exactly the part that came out wrong and run it again, instead of rewriting the whole prompt. Rewrite everything and you also change the parts that were working, and the next image goes wrong somewhere else — you end up going in circles. When a run uses credits, fewer wasted runs also means fewer credits spent. And what can happen to your words between you and the image model is a separate story: what can happen to your prompt before the image is drawn.

A face that belongs to no one

One last point belongs to the check itself. The face in the image belongs to a person who doesn't exist, so don't inspect it for who it resembles; that isn't a quality question. Where AI-generated models fit in a store's photos, and where they don't, is its own topic: AI fashion models for online stores.

With a little practice, the whole check takes under a minute: the full image as large as your screen allows, then at the size people will see, then a zoom into hands and the places where things meet, then a slow read of any text. That's enough to stop relying on a vague feeling and to know exactly what you're approving before it leaves your hands.

Frequently asked questions

How do I check an AI-generated image for mistakes?

Go in order of distance. Look at the whole image at full size and at the size people will see it, then zoom into hands and small objects, then the places where two things touch, then read any text letter by letter.

Why do AI images get hands wrong?

Hands are full of small, overlapping details, and Northwestern's research summary notes that AI models are better at faces than at hands: fingers can merge, lose their nails or come out at odd proportions. An extra finger alone doesn't prove anything about a photo, so treat hands as a place to check, not a verdict.

Why is the text in my AI image gibberish?

Because the letters are drawn as shapes, not typed as text, and unreadable text and logos are a known kind of error in AI-generated images. Change the prompt and run it again, or frame the shot so the text isn't in it.

Can I fix one detail without starting over?

Not inside Generate photorealistic images — it has no control that reshapes a finished image. Change the part of the prompt that produced the mistake and run it again; Generate again keeps your prompt for a fresh variation.

Do AI faces look real enough that people can't tell?

They can. In a 2022 study, people sorting faces into real and AI-synthesized scored 48.2% on average — about what guessing gets you — and 59.0% after training. The study tested the face generators of its time, but it's a good reason this check skips the face and goes to hands, edges and text.

Make one and check it

Generate a portrait with Generate photorealistic images, then give it this one-minute check before you download it. When a run uses credits, the cost appears in the editor before you run it.