Generate a character, like the result, then generate her again in a different pose. What comes back is usually a relative. Same hair, same jumper, same general vibe, and a face belonging to somebody else: the chin has changed, the eyes sit differently, and the whole read of the face is off by an amount you can’t name but spot instantly.
That gap has a name in the industry, character consistency, and it’s the hardest thing to buy in image generation today.
Consistent style and consistent character are two different things
Both get filed under the same word, which suits everyone selling a tool.
Style consistency means the images look like they came from the same pen: same line weight, same palette, same shading. Models are good at this. A style keyword, a reference sheet or a light fine-tune will hold a look across hundreds of images without much fuss.
Character consistency means the identity survives the pose, the expression, the angle, the lighting and the setting all changing at once. Far harder, and the two get conflated constantly in marketing copy, because a gallery of images in a matching style reads as consistent at a glance even when every face in it belongs to a different person.
Why a description can’t pin down a face
The reason is baked into how generation works.
A prompt is a description, and a description of a person is wildly underspecified. “Woman in her thirties, curly dark hair, round glasses, green jumper” matches a few billion faces. The model isn’t recalling a person, it’s sampling one from the enormous space that fits the words, and next time it samples again.
More words narrow it a little. They never close it, because faces differ along dimensions language handles badly: the distance between the eyes, how the cheek meets the jaw, what the mouth does at rest. There’s no vocabulary for any of that, and those are precisely the features people use to recognise each other.
Which is why prompting alone caps out well below what a reader needs. You can get the same type of person from a prompt. You can’t get the same person.
The three ways tools get closer
A prompt plus a fixed seed. Cheapest, and it holds only while nothing changes. Alter the pose or the scene and the face moves with it.
Reference conditioning. The model gets one or more images of the character alongside the prompt and is asked to keep that identity while changing everything else. This is what most consistent-character features are doing under the bonnet. It’s quick, it needs no training, and the quality depends heavily on how good the reference is and how far the new image strays from it.
Training the character in. A small fine-tune teaches the model the character as a concept. Strongest identity retention and much the most expensive, since it costs money and minutes per character, so any product doing it for every user has made a real commitment on its unit economics.
Most tools use one of the three. The interesting engineering is in what happens when it fails, because none of them holds an identity at every angle.
Readers notice the face and forgive everything else
Useful to know when you’re judging a tool: drift doesn’t cost you evenly.
Get the jumper wrong and nobody minds. Change the background, the hairstyle, even the body shape between panels and a reader absorbs it as artistic variation. Move the eyes fractionally further apart and the same reader, who couldn’t tell you what changed, reads the panel as a different character.
Face recognition is the thing human vision is best at and it runs before conscious attention gets involved. So a tool that keeps a costume pixel-perfect while the face wanders has spent its effort on the half of the picture nobody was checking.
Consistent doesn’t mean identical
A character that’s pixel-identical in every image reads as a sticker pasted into scenes. It looks dead. Faces need to move, and a real person does look different tired, laughing, in profile, in bad light.
The target is recognisability under variation: the character changes about as much as a real person would between two photographs, and no more. That’s a fuzzier target than pixel identity, and it’s why the honest test is a human glance rather than an image-similarity score. Two images can score well on any embedding metric and still read as cousins.
How to test a tool in five minutes
The demo gallery won’t tell you anything. Try these instead.
- Generate the character at three-quarter view, in profile, and head on. Profile is where most tools come apart.
- Put them in a panel with two other characters and watch whether the faces start borrowing features from each other.
- Ask for an extreme expression: crying, shouting, laughing with their eyes shut.
- Come back tomorrow, start a fresh session and generate them again from scratch. Holding a character inside one session is a much easier problem than holding it across sessions, and the second one is what a daily diary comic runs on.
- Do all of it ten times and lay the results out side by side. One generation proves nothing, since drift shows up as a spread rather than a single failure.
That last test is the one that matters, and no demo will ever run it in front of you.
Why we care about this more than the pictures
Stripkin’s premise is that the same illustrated you turns up strip after strip, week after week, for years. Everything else in the product is replaceable. If the character drifts, the archive stops being a record of a person and becomes a folder of unrelated AI pictures, and the internet is not short of those.
So consistency is the bar the product either clears or doesn’t, ahead of speed, cost, and how impressive any single image looks. It’s an odd thing to optimise for, and it’s the reason the strips are drawn in a simplified style rather than anything photorealistic: fewer identity-carrying features means fewer things that can quietly go wrong between Tuesday and Wednesday.