Where Nano Banana 2 sits
Nano Banana 2 is built on Google Gemini 3.1 Flash Image — the workhorse of the family. Native 4K output, support for up to 14 reference images simultaneously, and virtual try-on capability that stands out among comparable models.
On Nano Banana we run the Lite tier of this generation (gemini-3.1-flash-lite-image) in production; the standard 2 capabilities described here follow public information — verify current specs before you build a workflow around them.
What it's best for
- Virtual try-on: compositing clothing and accessories onto a model
- Multi-image fusion: combining product + scene + person into a single image
- 4K output: when you need high-resolution hero images or print-ready assets
- Multilingual text rendering: bilingual product labels, mixed Chinese/English posters
- High-volume batch production: the go-to model for daily e-commerce image output
Virtual try-on prompt
Reference image order: put the person image first, the clothing image second. The model defaults to compositing the person from image 1 with the clothing from image 2.
- Dress [the model in the provided / this person] in the provided [clothing type].
- Maintain [the model's face / hair / skin tone / body shape].
- [Background: clean white studio background / outdoor urban setting].
- The [clothing] should fit naturally with proper [folds / draping / tailoring].
Multi-image compositing prompt
- Combine these reference images: place the [product] in the given [scene],
- with [person/model] naturally interacting with it.
- Maintain consistent lighting, perspective, and color temperature throughout.
Text rendering tips
NB2 is more reliable for text than NB1, but still has limits: keep text to 8 characters or fewer per image, and add "clearly readable text, sharp letterforms" to improve accuracy. For longer text or brand-specific fonts, post-processing is more reliable.
2026 content safety changes
Since February 2026, NB2 has a much higher rejection rate for requests involving financial information (invoices, ID documents, price tags). This is a Google model-level safety policy that can't be bypassed. Handle these needs through other methods.