Nano Banana 2.1 tops its own image editing scores

Google has published a model card for Nano Banana 2.1, an image generation and editing model built on Gemini 3.6 Flash. The card, on deepmind.google, describes it as a member of the Gemini 3 series: it reads text and images as input and returns both an image and text. Input context runs to 1M tokens, image output to 4K tokens and text output to 64K tokens.

Multi-character consistency at 1106 Elo

The evaluation section reports side-by-side human preference scores, expressed as Elo, for Nano Banana 2.1 in a "thinking" and a "no thinking" configuration, against Gemini 3.1 Flash Image and Gemini 3 Pro Image. The thinking configuration leads every listed category.

Multi-character consistency is the highest figure on the page at 1106, against 978 for Gemini 3.1 Flash Image and 1011 for Gemini 3 Pro Image. Multi-reference editing reaches 1066, stylization 1062, mask and ink-based editing 1049, and infographic design 1048. Single-character consistency sits at 1028, general editing at 1026 and product consistency at 1024. The no-thinking configuration scores lower throughout, from 1068 on multi-character consistency down to 980 on general editing.

Infographic factuality is reported separately, on a different scale: 0.521 for the thinking configuration, 0.328 without thinking, 0.179 for Gemini 3.1 Flash Image and 0.265 for Gemini 3 Pro Image.

Where the card says it still falls short

The intended usage section lists editing tasks the model is aimed at: general image editing, stylization, character editing, object and environment editing, ink or doodle based editing, and multi-image work such as recontextualizing several products or characters in one scene. Google also names poster text, intricate diagrams, long context real-world knowledge and localized text rendering across languages.

The limitations section is specific. Small text often renders blurry at 1k, and long paragraphs and page-length text remain weak. Character consistency is not always perfect between input images and the generated output. Masked and doodle based editing shows partial instruction following and ink persistence, and the card notes rare cases where a subject's pose from the input image is retained in the output. Spatial localization, such as left and right, is occasionally confused. Google states the model is still limited in advanced capabilities with world knowledge, 3D reasoning and factuality, and that occasional slowness or timeout issues occur.

Distribution covers the Gemini app, Google AI Studio, the Gemini API, Google Search AI Mode, Google Ads, Google Flow and Google Stitch. The knowledge cutoff for the underlying Gemini 3.6 Flash is March 2026, with some domains limited to January 2025.