How many reference images can Nano Banana Pro take? The 14-image limit and how to prompt with several images
Nano Banana Pro accepts up to 14 reference images per request, with up to 5 people kept consistent and up to 6 objects reproduced at high fidelity. Limits for every model on Miku AI, how people actually use references, and how to write multi-image prompts.
Published · 5 min read
Short answer: 14. Google's documentation says Nano Banana Pro (Gemini 3 Pro Image) can mix up to 14 reference images in one request: up to 6 object images reproduced at high fidelity and up to 5 images of people kept consistent. On Miku AI, Nano Banana Pro, Nano Banana 2 and Nano Banana 2.1 all accept 14; once you reach 14 you cannot add more.
The Nano Banana Pro reference image limit
Nano Banana Pro is Google's Gemini 3 Pro Image model (API name gemini-3-pro-image-preview). According to the Google Gemini API documentation, its reference rules are:
| Item | Limit | What it covers |
|---|---|---|
| Reference images in total | 14 | Every reference image in one request |
| Objects (high fidelity) | 6 | Products, props, logos that must appear exactly as shown |
| People (consistent identity) | 5 | People whose faces must stay the same |
Google only promises fidelity and consistency within these numbers. If a product has to look exactly like the original, keep object images to 6 or fewer and people to 5 or fewer.
Reference image limits for each model on Miku AI
Limits in the Miku AI workspace as of 10 November 2026. The add button turns off at the limit. If you switch to a model with a lower limit, remove the extra images before you submit.
| Model | Reference images | Output |
|---|---|---|
| Nano Banana Pro | 14 | 1K / 2K / 4K |
| Nano Banana 2 | 14 | 1K / 2K / 4K |
| Nano Banana 2.1 | 14 | 1K / 2K / 4K |
| GPT Image 2 / 2.5 / 2.6 | 16 | 1K / 2K / 4K / 360° panorama |
| Qwen Image 2.1 | 10 | 1K |
| Seedream 5 Pro | 8 | 1K / 2K |
| Midjourney 8.2 | 7 | 1K |
| Grok Image 2.0 | 1 | 1K |
If you need more than 14 images in one request, use a GPT Image model. Within 14, Nano Banana Pro is the steadier choice for keeping faces consistent and products accurate.
How many references do people actually use?
We looked at Nano Banana Pro tasks on Miku AI over the 30 days up to 10 November 2026:
- About 89% of tasks included at least one reference image; plain text-to-image was only about one in ten.
- One image was by far the most common, about 64% of all tasks. Two or three images: about 20%. Four to six: about 5%. Seven or more: under 1%.
- The most references used in a single task was 12. Nobody has used all 14 yet.
- 2K was the most used output (about 44%), then 1K (about 30%) and 4K (about 26%).
- The median time from submission to finished image was about one to two minutes. Tasks with references take a little longer than text-only ones, but adding more references does not multiply the time.
So the 14-image limit is more than enough for nearly everyone. The common pattern is one reference plus a text edit; multi-image blends are mostly product composites and group shots.
How to write a prompt with several reference images
- Refer to images by upload order. References reach the model in the order you add them, so write "image 1", "image 2" instead of "the woman" or "the jacket".
- Say what each image is for. For example: "image 1 is the face, image 2 is the outfit, image 3 is only for colour and light". The model cannot tell which image is the subject and which is a style reference.
- At most 5 people and 6 must-match objects. That is the range Google promises fidelity for.
- Use clean references where the subject fills most of the frame. Cropping out clutter helps more than adding another image.
- Start small. Settle composition and style with two or three images, then add the rest. When something goes wrong it is easier to see which image caused it.
- Use 2K or 4K for busy scenes. With several objects in one frame, small items blur at 1K.
An example multi-image prompt:
Image 1 is a front-facing photo of the model, image 2 is a beige trench coat, image 3 is a brown leather tote, image 4 is a Paris street. Create a fashion magazine street-style photo: the model from image 1 wears the coat from image 2 and carries the bag from image 3, walking down the street from image 4, golden-hour side light, 35mm lens, the model filling two thirds of the frame. Keep her face and hair unchanged, and keep the colour, material and hardware of the coat and bag identical to the originals.
For ready-made prompts, browse Nano Banana Pro examples other users have shared in the prompt gallery and reuse one with a click.
FAQ
What happens if I add more than 14 reference images to Nano Banana Pro?
On Miku AI you cannot add a 15th image. If you need more material, combine several images into one collage first, or switch to a GPT Image model, which accepts 16.
Do reference images cost extra?
No. On Miku AI, Nano Banana Pro is charged per generated image and resolution (1K / 2K / 4K); one reference or fourteen costs the same. See pricing for the rates. Failed generations are refunded automatically.
Do Nano Banana 2 and Nano Banana Pro have the same reference limit?
On Miku AI both accept 14. Google describes the split between people and objects slightly differently for the two models, so check Google's documentation if you need the exact split. In general Nano Banana Pro keeps faces and fine detail more faithfully, while Nano Banana 2 is cheaper and faster: on Miku AI it usually takes half the time of Pro or less.
Do more reference images give better results?
Not necessarily. Every extra image adds a constraint the model has to balance, and the prompt has to explain what each image is for. About 64% of tasks use a single reference. Only add images whose content must appear in the result; one or two are enough for style.
Which formats and sizes work?
Miku AI accepts JPG, PNG and WebP. Use sharp originals where the subject is complete, not shrunken screenshots. For people, a front-facing or half-body photo in even light works best.