Google DeepMind · Gemini 3 Pro Image
Google's reasoning image model plans the layout before it renders. Use it to bring up to 5 people from separate photos into one shot, or to put your logo on a product in a finished scene, with up to 14 reference images and no Gemini subscription.
Multiple references
Yes. Upload one reference for each person or object and say what each one is for. In both examples the large image is the result and the small cards are the inputs.
Five people
Upload one plain, front-facing portrait per person, refer to each by appearance, say where each one goes and ask to keep faces, hair and clothes. Compare every face with its reference before you share the result.
Logo + product
Upload the logo and the blank product as separate references, then say which is which, where the logo goes and what must not change. Check the spelling and the product's shape in the result.
Start most new images with 2.1: Google recommends it for new projects, and it costs fewer credits here. Switch to Pro when one shot needs five people, when a logo and a product must match, or when you are continuing work made with Pro.
| Compare | Nano Banana Pro | Nano Banana 2.1 |
|---|---|---|
| Google model | Gemini 3 Pro ImageNovember 20, 2025 | Based on Gemini 3.6 FlashOctober 6, 2026 |
| How Google positions it | Premium choice for complex tasks | Recommended for new projects |
| People kept consistent (Google) | Up to 5 | Up to 4 |
| Objects kept at high fidelity (Google) | Up to 6 | Up to 10 |
| Human-preference testsDeepMind 2.1 model card, Oct 2026 | Lower in 9 of 9Infographic design: 912 Elo | Higher in 9 of 9Infographic design: 1048 Elo |
| Credits per image | 1K / 2K: 30 · 4K: 60 | 1K: 8 · 2K: 12 · 4K: 27 |
| 5 portraits into one 2K image | 30 credits | 14 credits |
| Best for | Five-person shots, brand assets | Most new images |
Explainers
Yes: write the facts and every label into the prompt and describe the layout. Pro reasons through the arrangement, but it doesn't check your numbers, so proofread the result. For a stand-alone explainer, Nano Banana 2.1 is usually the better pick: Google's model card scores it higher on infographic design, and it costs fewer credits here. Choose Pro when the diagram has to match earlier Pro work, or when the same image also needs several people and a logo.
Data infographic
Supply every name and value yourself, then set the sort order, a shared baseline and where each label sits, so the shapes stay in proportion, for example peaks ranked by height. Afterwards, check that the largest shape carries the largest number.
Labeled cutaway
Name each layer with its range and ask for leader lines, so every label points at its own band instead of floating nearby. Check that the labels run in the same order as the layers.
Pro reads full sentences and reasons through the request, so specific details matter more than keywords. These habits come from Google's prompting guides and from problems developers report.
Brief it like a designer
Name the subject, layout, style and where the image will run. The intended use tells Pro how large the text should be and how much space to leave.
Supply the facts and the exact words
Put every headline, label and number in quotes, and render at 2K so small type stays sharp. On Oeneo AI, 2K costs the same as 1K for Pro.
One clear portrait per person
Use a front-facing photo with the whole face visible, and describe each person by appearance so the prompt can say who goes where.
One job for every other reference
Say what each image is for — for example, "image 1 is the logo; print it on the label of the bottle in image 2" — and what may change.
Paragraphs of text
Long paragraphs and page-length text still break. Keep the words in the image short and add body copy in a layout tool.
Editing the same copy again and again
Details drift a little with each round. If faces or textures start to slip, go back to the original image.
A list of what you don't want
Describe what should be there: "an empty beach at sunrise" works better than "no people, no boats".
Publishing without a proofread
Check every number, label, logo and face against your sources before the image goes out.
Example edit prompt: "Using the provided photo, change only the lighting to warm afternoon sun from the left. Keep the shop front, the sign, the people and the camera angle exactly the same."
Each model has its own page and runs in the same workbench on one credit balance.
New
GoogleGoogle's recommended model for new projects, at fewer credits per image than Pro.
View model
GoogleThe previous Flash generation, for continuing an existing series or reusing prompts tuned for it.
View model
GoogleQuick text-only drafts when you don't need reference images.
View model
New
Black Forest LabsComposed scenes from a prompt or references
View modelRemove the background, upscale, vectorize or denoise any result, all on the same credit balance.
Transparent PNG or WebP with clean hair and product edges, ready for a listing or a new composition.
2× / 4× / 8× with recovered texture instead of smeared pixels, so a 1K draft can become a print.
Logos, icons and flat illustrations become editable SVG that scales to any size.
Remove grain and compression artifacts from low-light photos and scans while keeping real detail.
Nano Banana Pro is Google DeepMind's image generation and editing model; its official name is Gemini 3 Pro Image. Google released it on November 20, 2025, built on Gemini 3 Pro, and it reasons through each prompt before rendering. There is no Nano Banana Pro 2: the models released after it are the faster, Flash-based Nano Banana 2 and Nano Banana 2.1.
For most new images, start with Nano Banana 2.1. Google's API documentation recommends 2.1 for new projects, and the 2.1 model card from October 2026 scores 2.1 above Pro in every human-preference test it lists, including infographic design and multi-character consistency. On Oeneo AI, credits per image are 1K: 8 · 2K: 12 · 4K: 27 for 2.1 and 1K / 2K: 30 · 4K: 60 for Pro. Pro still earns its place when one shot needs five people (Google documents up to 5 characters for Pro and 4 for 2.1), when a brand's logo and product must match, or when you are continuing work made with Pro.
Nano Banana 2 is the previous Flash generation: use it to continue an existing series or to reuse prompts you tuned for it (credits per image: 1K: 16 · 2K: 24 · 4K: 32). Nano Banana 2 Lite takes text only and doesn't accept reference images; it is the quickest way to sketch ideas (credits per image: 1K: 8).
You don't need a Gemini subscription, but Pro isn't free here: sign in, choose Nano Banana Pro and pay per image, and failed tasks are refunded automatically. New accounts get 10 free credits — not enough for a Pro image (30 credits), but enough for a 1K image with Nano Banana 2.1 or 2 Lite (8 credits). Images made on the free plan are public; paid plans keep them private.
Up to 14 per run. Google's documentation says Pro keeps up to 5 people consistent and up to 6 objects at high fidelity within that total; use the rest for style or setting. Give each person one clear, front-facing photo, describe them by appearance in the prompt, and compare every face with the original — resemblance is strong but not guaranteed.
Small text is a known weak spot: Google's model card says text is still poor at small sizes, especially at 1K, and in long paragraphs. Keep the words short and in quotes, give them room in the layout, and render at 2K or 4K. Simplified Chinese is one of Google's best-supported languages, but tiny or dense characters can still break, so ask for a bold sans-serif and proofread every character.
No. Google offers Search grounding for Nano Banana Pro, but it is turned off on Oeneo AI, and the model's own knowledge ends in January 2025. For prices, scores, schedules or recent events, write the facts into your prompt and check them in the result.
It happens occasionally. Developers report that the model sometimes ignores a reference, and Google's model card notes that consistency isn't perfect and that edits can carry over parts of the input. Name each reference by what it shows, say what must stay the same, change one thing per run, and generate again if a reference was skipped. If quality slips after several rounds, start from the original image.
Use 2K for most screens and for anything with small text, which is often blurry at 1K. Use 4K (60 credits per image) for print or large hero images. On Oeneo AI, 2K costs the same as 1K for Pro.
Yes, on paid plans: subscriptions include a full commercial license, and the Lifetime plan covers personal commercial use (see Pricing). Whether a specific image is safe to publish still depends on what it shows, such as real people, logos and brands, and on the rights in your reference images. Google also embeds an invisible SynthID watermark in every Nano Banana Pro image to mark it as AI-generated; it doesn't change how the image looks.
Add your references, quote the exact words and render at 2K. The Generate button shows the credits before you submit.
