Blog

Character consistency in an AI image generator: what holds and what vanishes

One master portrait and the Character role keep a face, hair, age and wardrobe across generations. Scars, moles and tattoos do not survive. The sentence that makes a reference stick, and four habits that fix the rest.

September 20, 20265 min readAIKIZI

To keep one character's face across many AI images you need a clean master portrait and a sentence that tells the model the reference is that person. On aikizi that sentence is written for you by the Character role. It carries the face, the hair, the age and the clothes. It drops scars, moles and tattoos — and knowing that in advance is the difference between a usable set and forty near-misses.

The master, and one move

We made a character for this page and moved him to a new place, a new posture and a new light for 8 credits, with one reference attached.

A studio portrait of a man in his sixties with a full white beard, heavy grey brows and a small mole below his right eye, in an olive linen kurta against plain grey — The master, made for this page. Nano Banana Pro, 1K, 4:5 — **8 credits**. Plain backdrop, flat key from the left, nothing happening. Its decode.

The same man seated cross-legged in a workshop, leaning over a half-carved wooden puppet, both hands working, wood shavings on the bench, warm window light from the right — Same man, new place, new light, new posture. **8 credits**, one reference, the Character role. Its decode.

What held, and what did not

Feature	Survived?
Face structure — nose, mouth, jaw, eyes	Yes, without effort
Hair: length, texture, how it sits	Yes
Age and build	Yes
Wardrobe — the olive linen kurta, its collar and cut	Yes
Hands	Yes, both correct
The small mole below his right eye	No — gone entirely

Read that table as a buying guide rather than a curiosity. What makes your character recognisable at a glance — the shape of the face, the hair, the build, the clothes — is what these tools are reliably good at, and it is why a consistent character is achievable at all. What makes them specific is not covered, and you carry it yourself, prompt by prompt. Pick a character whose identity lives in the first list, or budget for the second. Part 6 explains why the split falls where it does.

The sentence that does the work

Gemini has no typed reference slots. Every image goes into one flat array with the prompt appended last, and nothing in the API says "this one is the face" — the model works it out from the words alone. So aikizi writes the words. Tap the Character badge on a thumbnail and this goes above your prompt, with the reference's position number in it:

Reference 1 is the character — keep their exact face, features, hair and age. Do not replace them with anyone else.

There are six roles in all — character, style, object, scene, pose, lighting — plus a custom one where you write the sentence yourself. Using the wrong one is the commonest reason a reference "does not work": the Style role explicitly says ignore its subject and composition, so a face attached as Style will not survive.

Four habits that fix most of it

Make one master portrait on a plain background, in flat even light, head and shoulders. Reuse it forever; do not make a new master per shot.
Attach it with the Character role on every generation that person appears in.
Name every mark in that generation's prompt — "the small mole below his right eye is visible" — and say which side. Do not rely on the reference alone.
Check the marks at 1K for 8 credits before you go bigger, then upscale for 15. Regenerating at 4K restarts the whole gamble.

What this is not

Nothing here trains a model on your character. There is no fine-tune, no LoRA, no stored identity: every generation re-derives the face from whatever you attach that time. That is a real limit and also the reason there is nothing to set up — you can have a consistent character five minutes after making the master. If you need many variations at once rather than one directed shot, a SUBJECT realm takes 1 to 5 references and produces poses, expressions and turnarounds from the same pattern.

The honest plan note: Basic allows 3 reference images per generation and Pro allows 10. For character work that gap matters more than the credit count, because a mark that appears in three references is much harder for the model to drop than one that appears in a single reference.

FAQ

Which AI image generator keeps a character consistent?

Any generator that lets you attach a reference image and say what it is. On aikizi that is the Character role, which writes an explicit keep-this-face instruction into the prompt. Where a tool gives you no way to label the reference, the model guesses its job — which is why those tools feel unreliable rather than wrong.

What makes a good master reference picture?

A boring one. Plain background, flat even light, head and shoulders, neutral expression. The less the picture is doing, the more of it is face, and the face is the only part you want travelling into the next generation.

How many reference images do I need?

One clean master is enough for the face. More helps for the details — a mark appearing in three references is much harder to drop. Basic allows 3 per generation, Pro allows 10.

Why does adding more references make my character worse?

Because every unlabelled image is another vote on an average, so ten similar references blur a face rather than sharpen it. Add references only when they do genuinely different jobs, and label each one.

How long does it take to set a consistent character up?

About as long as one generation. There is nothing to train and nothing to configure — make the master, then attach it. A master plus one re-pose is 16 credits and a few minutes.

Make one plain portrait in the Studio, give it the Character badge and re-pose it once — the whole technique is in that round trip. The longer version, with the full failure list, is part 6 of the generation series; the other five reference roles are in part 4.