You have a scene in mind: five recurring characters, a specific setting, a prop or two, and a color palette you have already locked in from earlier work. You have gathered the reference material, fourteen images in total, and now you want a single generated frame that pulls all of it together without any of your characters drifting into a stranger's face. Anyone who has tried this by hand knows the hard part is not the composition. It is keeping every character recognizable when the model is juggling that many inputs at once.
This guide walks through how the Nano Banana family handles multi-image reference composition, how to structure your references and prompt so five distinct characters stay consistent, and which tier to reach for depending on whether you are optimizing for raw quality or for the specific 14-image workflow.
What multi-image reference composition actually does
Most image models let you supply one reference and nudge the output toward it. Multi-image reference composition is a step up: you supply several images at once, and the model treats each as a source of visual information it can draw from when building a new frame. One image might contribute a character's face, another a costume, another the lighting of a room, another the shape of a prop.
The value for a multi-character scene is obvious. Instead of describing five faces in words and hoping the model invents something close, you hand it the actual reference for each character. The model has direct visual anchors to work from, which is what makes consistency possible in the first place.
Within the Nano Banana family on Atlas Cloud, this capability is documented on [Nano Banana 2 Lite](https://www.atlascloud.ai/models/nanobanana-2), which supports up to 14 reference images plus multi-image composition across 14 aspect ratios, with sub-2-second latency. That is the tier whose feature set maps directly onto a "combine 14 references" task. Nano Banana Pro is the higher-quality line, built for 1K, 2K, and 4K output when the finish of the frame matters more than raw input count. We will cover how to choose between them below.
Character consistency is a tagging and description problem
Handing the model fourteen images is only half the job. If you drop five character references into a request with no structure, the model has no reliable way to know which face belongs to which character in your scene, and that is exactly where identities blur or swap.
The fix is to treat each character as a labeled entity, not an anonymous input. Three techniques do most of the work:
- Per-character reference tagging. Give every character a stable name or label in your prompt, and associate each label with its reference image. Instead of "five people in a cafe," you describe "Mara (reference 1), Devon (reference 2), Priya (reference 3), Ari (reference 4), and Kaito (reference 5) seated at a corner table." The named anchor tells the model which visual source maps to which role in the scene.
- Consistent prompt descriptors. Keep the same distinguishing descriptors for each character every time you mention them: hair, build, clothing, a signature accessory. If Mara has "short silver hair and a green scarf" in one prompt, she keeps exactly those words in the next. Reusing the descriptor language across a series is what lets a character survive from frame to frame.
- Edit and reference-to-image mode. When you already have a good version of a character or a scene, use reference-to-image or edit mode rather than starting from a blank text prompt. Feeding the model your prior output as a reference locks in the look you already achieved instead of asking it to reinvent the character.
None of this depends on a secret parameter. It is disciplined structure: name your characters, anchor each name to a reference, and never let the descriptor language drift.
Key technique steps for a 14-image, 5-character frame
Here is a repeatable order of operations that keeps the process manageable.
- Sort your fourteen references by role before you write anything. Group them: five are character faces, the rest are setting, wardrobe, props, and palette. Knowing what each image contributes stops you from describing them all as interchangeable.
- Assign a stable label to each of the five characters and write a one-line descriptor for each that you will reuse verbatim across every generation.
- Write the composition prompt so it references characters by their labels and places them in the scene explicitly ("left to right," "in the foreground," "behind the counter"). Spatial instructions reduce the chance two characters merge.
- Attach the references in the request and describe what each group of references is for, so the model knows a given image is a face to preserve versus a room to borrow lighting from.
- Generate, then inspect each of the five faces individually. Consistency problems almost always show up in one or two characters, not all five at once.
- For any character that drifted, run an edit or reference-to-image pass on just that region or that character, feeding the correct reference again, rather than regenerating the whole frame from scratch.
Because the exact request shape (how references are attached, how many fields, what each is named) belongs in the API spec and can change, confirm the current structure at atlascloud.ai/docs under the image models section rather than hard-coding assumptions. The technique above holds regardless of the field names.
Doing it on Atlas Cloud
Atlas Cloud is a full-modal AI inference platform that curates 300+ SOTA models across text, image, and video behind one OpenAI-compatible endpoint. The whole Nano Banana family lives on that same endpoint, reachable with one API key and one billing account, which matters here because a multi-character project tends to bounce between tiers as you iterate.
For this specific task you have two sensible tiers:
- Nano Banana 2 Lite is the efficiency-focused tier that explicitly supports up to 14 reference images, multi-image composition, and 14 aspect ratios, at $0.04 per image (Developer tier drops to $0.028, a 30% reduction). Its sub-2-second latency makes it the natural choice for the iterative loop this workflow demands, where you generate, check five faces, fix one, and generate again. When your task is literally "combine 14 references," this is the tier whose documented feature set fits.
- Nano Banana Pro is the higher-quality Pro line (Google's Gemini 3 Image Pro family) with 1K, 2K, and 4K output. Standard text-to-image and edit are $0.14 per image, the Ultra text-to-image and Edit Ultra variants are $0.15, and the Developer tier halves the standard rate to $0.07. Reach for Pro when the final frame needs to be delivery-grade at high resolution and you are willing to trade the Lite tier's specific 14-image convenience for finish quality.
A practical pattern is to compose and iterate on the Lite tier, where the reference-image workflow and low latency make the trial-and-error cheap, then produce the final locked frame on Pro at the resolution you need. Every model shows its live price next to the Run button in the Playground, so you confirm the exact per-image cost before writing any code, and the full catalog is browsable at atlascloud.ai/models. Because the endpoint is OpenAI-compatible, an app already built on the OpenAI SDK reaches these models by changing the base_url and API key, with no rewrite.
Tips for keeping five characters consistent
- Lock your descriptor language early. Write the five one-line character descriptions once, save them, and paste them unchanged into every prompt. Rewording a character mid-project is the most common cause of drift.
- Keep the highest-quality reference for each face. A clean, well-lit, front-facing reference gives the model far more to anchor to than a blurry crop, and it pays off across every frame that character appears in.
- Reduce competition in a single frame. Five characters plus fourteen references is a lot to balance. If two characters keep blending, generate them in a tighter grouping or split the scene and composite, rather than forcing all five into one crowded pass.
- Reuse your best output as a reference. Once a character looks right, feed that frame back in via reference-to-image mode so later generations inherit the approved look instead of re-rolling it.
- Fix locally, not globally. When one face slips, edit that character rather than regenerating the entire composition, which protects the four characters that already came out correctly.
FAQ
Q: Which Nano Banana tier actually supports 14 reference images? A: Nano Banana 2 Lite is the tier documented to support up to 14 reference images plus multi-image composition, at $0.04 per image. Nano Banana Pro is the higher-quality 1K/2K/4K line at $0.14 to $0.15 per image, best for a delivery-grade final frame.
Q: How do I stop the model from swapping my characters' faces? A: Give each character a stable label and a fixed one-line descriptor, anchor each label to its reference image, and reuse that exact language in every prompt. Named, consistently described characters are far less likely to blur into each other.
Q: Do I need a special API parameter to attach references? A: The technique is conceptual: name your characters, tag each to a reference, and use edit or reference-to-image mode for known looks. For the exact request shape and field names, check the image models documentation at atlascloud.ai/docs, since that is the authoritative source.
Q: Can I use both tiers in one project without separate accounts? A: Yes. Both Nano Banana 2 Lite and Nano Banana Pro sit on the same Atlas Cloud endpoint, so one API key and one billing account cover iterating on Lite and finishing on Pro.
Q: What if two characters keep merging in a crowded scene? A: Reduce the load in a single pass. Use explicit spatial placement in the prompt, generate the two problem characters in a clearer grouping, or split the scene and composite the results rather than forcing all five into one dense frame.
The bottom line
Combining fourteen references while holding five characters consistent is less about a hidden setting and more about structure: sort your references by role, give each character a stable label and a descriptor you reuse verbatim, anchor every label to its reference, and lean on edit or reference-to-image mode to lock in looks you have already achieved. On Atlas Cloud, Nano Banana 2 Lite is the tier built for the 14-image, multi-image composition workflow at $0.04 per image, while Nano Banana Pro delivers the high-resolution finish at $0.14 to $0.15, both on one OpenAI-compatible key. Iterate cheaply on Lite, finish on Pro, and confirm the exact request shape in the docs before you build.
