Blog
Telling an AI what each reference picture is for
Character, style, object, scene, pose and lighting — the six roles, the exact sentence each writes into your prompt, and why an unlabelled reference behaves unpredictably.
When you attach a picture to a prompt, the model does not know what job that picture has. It guesses from your words. Naming the job — character, style, object, scene, pose or lighting — is the single change that makes references behave, and on aikizi picking one writes an explicit sentence into the prompt for you. Here is what each role says and what each one is for.
Why a reference needs a job at all
Gemini has no typed reference slots. Every image you attach is pushed into one flat array and your prompt text is appended last. Nothing in the API says "this one is the face" or "this one is only the look". The model works it out from the words alone — so the only way to make a role stick is to say it.
That is why the same uploaded picture can behave in opposite ways. Attach a portrait and write "a woman at a workbench" and you are asking for a woman; you will get one, and it will not be yours. Attach the same portrait under the Character role and the model is told, in as many words, to keep that exact face.
The six roles, and the sentence each writes
| Role | Use it for | The exact sentence it adds |
|---|---|---|
| Character | A person whose face must survive | "Reference n is the character — keep their exact face, features, hair and age. Do not replace them with anyone else." |
| Style | The look only, not the subject | "Reference n is a style reference only — take its palette, mood and rendering quality, and ignore its subject and composition." |
| Object | A thing that must stay the same thing | "Reference n is an object — reproduce it exactly: its shape, proportions, markings and materials." |
| Scene | The place and the framing | "Reference n is the scene — keep its location, composition, camera angle and depth." |
| Pose | Body position and gesture | "Reference n is a pose reference — copy its body position, limb placement and gesture, nothing else." |
| Lighting | Where the light comes from | "Reference n is a lighting reference — copy its light direction, quality, contrast and colour temperature, nothing else." |
| Custom | Anything else | Your own words — "Reference n is the second character." |
The n is the reference's position in the list, and it matters: the number is the only handle the model has on a particular picture. An unlabelled reference is skipped rather than guessed at, because a wrong sentence is worse than none.
Two roles, same picture, opposite results
The Style clause is the one worth reading twice, because of what it explicitly forbids: ignore its subject and composition. Here is that instruction doing its job. Three pictures of a brass temple lamp were attached to a prompt about a clay pot, badged Style.
The reference was a photograph of a glass of water on concrete — nothing to do with kettles or rooms. What crossed over is exactly what the clause names and nothing else: the hard single source from the left, the high contrast, the long cast shadow, the cold colour temperature. The kettle stayed a kettle and the table stayed a table. That precision is what a role buys you — the same picture badged Style would have brought its concrete and its glass along too.
Choosing the right one
| What you want to keep | Role | Common mistake |
|---|---|---|
| This person's face | Character | Using Style, which tells it to ignore the subject |
| This product, exactly | Object | Using Style, which gives you a lookalike |
| This mood and palette | Style | Leaving it unlabelled, so the subject leaks in |
| This room, this camera angle | Scene | Using Style, which drops the composition on purpose |
| This body position | Pose | Expecting the clothes to come too — they do not |
| This light | Lighting | Describing the light in words as well, and contradicting yourself |
How many, and what they cost
References are free: a generation costs 8 credits at 1K on Nano Banana Pro whether it carries none or ten. What is limited is the count — 3 per generation on the $8 plan, 10 on the $20 one. Ten unlabelled references are worse than three chosen ones, because every image is another vote on an average. Ten is worth having when each does a different job: this face, in this place, under this light, holding this object. That combination is the real argument for the higher count, and it is the one a feature table never explains.
FAQ
What do reference image roles do?
They tell the model what job each attached picture has, by writing an explicit sentence into the prompt — keep this face, take only this palette, reproduce this object exactly. Without a role the model guesses from your wording, which is why unlabelled references behave inconsistently.
What is the difference between a character reference and a style reference?
Opposite instructions. Character says keep the exact face, features, hair and age. Style says take the palette, mood and rendering and ignore the subject and composition — so a face attached as Style will not survive.
Which role keeps a product identical?
Object: "reproduce it exactly — its shape, proportions, markings and materials". Style would give you something that resembles your product rather than your product.
Does a pose reference bring the clothes with it?
No. The Pose clause copies body position, limb placement and gesture and says "nothing else". Wardrobe travels on the Character role or in your own words.
How many references can I attach?
Three per generation on the $8 plan and ten on the $20 plan. They cost nothing extra, but ten similar ones blur the result — use the higher count when each reference does a genuinely different job.
Attach one picture in the Studio and change its badge from Style to Character on the same prompt; the two results explain the whole system faster than any table. The longer version is part 4 of the generation series.