The first time I used the full reference allowance, the image got worse.
Fourteen images: the character, two angles of her face, the jacket, the studio, the palette, a texture, a light reference, three mood boards. What came back was a competently lit photograph of nobody in particular. The model had done what a model does with fourteen inputs: it found the average.
That is the thing to understand about reference budgets. The number is a ceiling, not a target, and every image you attach past the useful set is a competing instruction.
I ran into this on Callirra, which exposes fourteen slots and no opinion about how to use them.
What the model actually accepts
| Limit | Value |
|---|---|
| References per run | 14 |
| Character references documented by Google | 4 |
| High-fidelity object references documented by Google | 10 |
| Prompt length | 20,000 characters |
| Output resolutions | 1K, 2K, 4K |
| Credits per image | 3 / 4 / 5 |
The 4 + 10 split is the most useful number on that table, because it tells you what the allowance is for: up to four people and up to ten objects, not fourteen of anything.
The four roles, not fourteen images
A reference set is a cast list. Each image has one job, and if two images have the same job, one of them is noise:
- Identity — the face or the product. One clean, evenly lit, front-on image. This one is non-negotiable.
- Structure — the angle or the pose you intend to render. Two or three quarters of the same subject, from the same reference shoot if possible.
- Material — the thing that has to look right at 100%: fabric, glass, condensation, powder, skin texture.
- World — palette, light direction, background. This is the reference people over-spend on, and it is the one that survives best as a written sentence instead of an image.
Four images cover all four roles for a single character. Ten covers a product range. Fourteen is for a scene with four characters and several objects — which is exactly the case the 4 + 10 split describes.
Spend the budget like this
| Shot | Identity | Structure | Material | World | Total |
|---|---|---|---|---|---|
| Portrait with a wardrobe change | 1 | 2 | 1 | 0 | 4 |
| Product hero on a set | 1 object | 1 angle | 1 | 0 | 3 |
| Line-up of four characters | 4 | 4 | 2 | 1 | 11 |
| Editorial still life, two products | 2 objects | 1 | 2 | 1 | 6 |
Note what is missing from every row: mood boards. A mood board is a description of taste, and taste is cheaper to write than to reference.
Say what each reference owns
The prompt does not have to describe the images — that is the reference's job. It has to say which image owns which decision:
Use the supplied images as follows:
- image 1 owns identity: face, hair, skin tone. Do not change her features.
- image 2 owns the jacket: cobalt wool, exact cut.
- image 3 owns the light: single soft key from camera left, deep shadow right.
- image 4 owns the palette only.
Render her in a sunlit gallery, three-quarter view, same identity and jacket.Compare that with "use the references", which is what produces the average. The moment a reference has an owner and a scope, conflicting inputs stop being a vote and start being a spec.
Where it goes wrong
Two references that disagree. Two jacket photos in different blues means the model picks a third blue. Pick one and delete the other.
References that fight the prompt. "Preserve the exact shape and colour" alongside three shape variants is a request to reconcile the irreconcilable.
Assuming PNG means transparency. It does not on this model: choosing PNG does not guarantee a transparent background, and there are no background or quality controls here at all — aspect ratio, resolution, JPG/PNG and references are the whole surface.
Treating 14 as headroom to use "just in case". Every unused reference is a liability: it is one more thing the model must decide the weight of.
The iteration trick that beats a bigger reference set
Reuse the output as the next reference. The model keeps identity better across a chain of edits than across a pile of inputs, and the chain is auditable: each step has one changed instruction.
- Render the base frame with four references.
- Feed that frame back, change the background, keep identity locked in the prompt.
- Feed the result back for the wardrobe change.
Three steps, three credits each at 1K, and a character who stays the same person — which is what people actually mean when they ask for "consistency".
FAQ
Does more references cost more credits? No. The price is per resolution: 3 / 4 / 5 for 1K / 2K / 4K.
Can I mix characters and objects freely? Within the documented split — up to four characters and up to ten objects.
Is a reference image required? No, text-to-image works. References are what you add when identity or material has to hold.
What if the text on a product must be exact? This model is strong at text, but review it at 100% anyway; small type is where every model still slips.
The reference limits and the 4 + 10 split are documented on Callirra's Nano Banana 2.1 page. If you have a fourteen-reference workflow that consistently beats a four-reference one, I want to see the prompt — that would break the pattern above.