Guide
Why AI changes the face every time
You wrote the same prompt. You got a different person. That is not a bug you are hitting, and no amount of extra description fixes it — it is what the model is doing.
The short answer
A prompt names a category, not a person
An image model starts from random noise and removes it, step by step, until what is left matches your words. The words steer that process. They do not identify anybody.
“A woman in her late twenties with dark hair and green eyes” is a true description of millions of different people. The model is not choosing between them, because it does not hold a list of people to choose from. It is producing something that satisfies the description, and the noise it started from decides which of the millions you get.
Run it again and the noise is different. So is the face. Nothing carried over, because there was nothing to carry — the model keeps no memory of the last image it made, and would not know which parts of it you cared about if it did.
The usual first fix
Why adding more adjectives stops working
The instinct is to describe harder. A narrower description does narrow the range — “a small scar through the left eyebrow” rules out everyone without one — and for the first few details it genuinely helps.
It stops helping quickly, for two reasons. The first is arithmetic: each detail divides a vast space into a slightly less vast one, and you run out of describable features long before you reach one person. The second is that models weight a prompt unevenly. Past roughly a dozen clauses, later ones are quietly diluted, and you can find yourself adding detail that changes nothing at all.
There is also a failure that looks like success. Pile on enough unusual features and every output starts to look the same — because you have described something so specific that the model has few ways to render it. That is not your character. It is a corner of the model, and it will not survive being put in a new pose.
The second fix
Seeds give you repeatability, not control
The seed is the number that decides the starting noise. Hold the seed, the prompt, the model and every setting identical and you get the identical image back. That is genuinely useful, and it is why seeds get recommended so often.
It is also the wrong tool for this problem, because the thing you want is the same person in a different picture. The moment you change the prompt to put her somewhere else, the denoising path changes from the first step, and the face goes with it. You are not moving a character through a scene. You are making a new image that happens to start from the same noise.
Small prompt edits sometimes produce a face that is close enough to pass, which is where the advice comes from. It is luck, it fails without warning, and it fails completely as soon as the camera angle moves.
What actually works
Stop describing the face and show it one
Every approach that holds a character has the same shape: the identity stops being carried by words and starts being carried by something the model can look at or has been taught. A reference image the generation conditions on. An embedding trained to mean this face. Weights fine-tuned on a set of photographs.
They differ in cost, in how much they hold, and in how they break. Those differences are the whole decision, and the next guide walks through them one at a time — including the cases where training a LoRA is still the better answer.
One thing worth knowing before you get there: a single reference image gets you a long way and then hits a wall at the first profile shot. A model conditioned on a front-facing photograph has never been shown the side of that head, and will invent it. That is the problem the next two guides spend most of their time on.
Keep reading
The rest of the series
- How to get the same character in every imageFive approaches, what each costs, and the cases where training a LoRA is still the right answer.
- Keeping a character consistent in AI videoWhy video drifts twice over, and the two techniques that fight it from opposite directions.
- Why the outfit and the room change tooWardrobe and location drift for a different reason than faces do, and need a different fix.
Where this comes from
Unlimited Studio is built around this one problem
Everything above is what the product does: a character is locked from eight reference angles, each generation conditions on whichever angle its prompt implies, a score checks the result against the reference, and a film is shot one frame-chained shot at a time. What it will and will not generate is set out in the content policy.