Nano Banana 2.1: Image Generation and Editing for Product Workflows

Nano Banana 2.1 is now available on RouterHub as google/gemini-nano-banana-2.1. Google’s image model brings improvements in visual quality, prompt adherence, text rendering, and consistency across edits. It accepts text and image inputs and returns image and text outputs, making it relevant to applications that need both original visuals and controlled revisions.

For developers, the useful unit of work is often a sequence: start from a product reference, generate a scene, change its composition, and adapt the result for another placement. Nano Banana 2.1 focuses on the generation and editing capabilities that support that sequence.

More control over text and composition

An image can look convincing and still fail its brief if the headline is wrong, a label is missing, or the layout hides the product. Google highlights improved text rendering and prompt adherence in Nano Banana 2.1, including better handling of text and infographic layouts.

That matters for visual features such as localized promotional images, product explainers, and diagrams. A developer can supply the intended wording, hierarchy, and composition as part of the request, then review whether the result preserves those requirements. Text inside the image remains an output to check, particularly when it is small, lengthy, or essential to understanding the visual.

Editing around a reference

Nano Banana 2.1 supports conversational editing: an existing image can become the starting point for a new instruction. Google reports improvements in general editing, product consistency, and mask- or ink-guided changes. These capabilities are useful when the goal is to revise part of an image while retaining its recognizable subject.

Consider a product-visual application. A user supplies a reference, requests a new background, and then asks for a different arrangement. The application needs to preserve the product’s shape, markings, and colors throughout those revisions. Product consistency is therefore a separate quality dimension from whether the new scene looks attractive.

For localized edits, a mask or drawn instruction can make the requested area more explicit where the integration supports those controls. The resulting image still needs review for changes outside that area and for details that drift during revision.

What Google’s image evaluations show

Google’s October 2026 side-by-side human evaluation reports these Elo scores for the models with Thinking enabled:

Evaluation	Nano Banana 2.1 (Thinking)	Nano Banana 2 (Thinking)
Overall text-to-image preference	1050 ±14	990 ±7
General editing	1026 ±12	938 ±11
Product consistency	1024 ±18	955 ±22

These are human-preference Elo scores with reported uncertainty ranges, not accuracy percentages or RouterHub-run results. The improvements span both initial generation and revision, which makes the comparison relevant to applications that keep working with an image after the first output. The scores do not guarantee the same preference on every product, style, or editing instruction.

Formats for different visual placements

Google documents 1K, 2K, and 4K image output, with 1K as the default, along with wide and panoramic aspect ratios including 4:1 and 8:1. Those model capabilities create options for banners, extended compositions, and higher-resolution assets; the controls available to an application depend on its API integration.

A practical workflow starts with the intended placement. A wide website banner needs space for its subject and any accompanying page text. A product detail image needs a clear view of the object. Specifying the composition at generation time can reduce the need to force one image into every placement later.

Nano Banana 2.1 on RouterHub

Detail	Nano Banana 2.1
Provider	Google
RouterHub model ID	google/gemini-nano-banana-2.1
Inputs listed on RouterHub	Text and images
Outputs	Images and text
Context window listed on RouterHub	131,072 tokens
Primary uses	Image generation and iterative image editing

From a generated image to a usable asset

The strongest starting point is a bounded workflow with recognizable acceptance criteria: preserve the supplied product, reproduce the required wording, follow the requested composition, and make the requested edit. Keep the original reference available through the sequence so reviewers can compare identity and detail across versions.

Character consistency remains imperfect, and spatial or masked-edit instructions can fail. Build review around the details that matter to the finished asset. Nano Banana 2.1 gives developers a model for generating, revising, and adapting visuals within one image workflow, with improvements that extend beyond the first render.