Nano Banana 2.1: ‘Just change the light.’ Is AI finally listening?

A small version number, a useful promise

Google describes 2.1 as an update to Nano Banana 2 for image generation and conversational editing. It claims better instruction following, character consistency across successive edits and text rendering. Output supports 1K, 2K and 4K; up to 14 reference images can inform one request, with consistency support for up to four characters and ten objects. Google also reports fixes for tiling artefacts in very wide 2K and 4K panoramas.

Those are useful promises for a director assembling a treatment or a designer adapting a campaign. A reference gives the model something specific to preserve. Several references can describe a character, a location and a product together. The creative question becomes how faithfully the result respects those decisions when the setting changes.

For me, the important distinction is between making another attractive image and continuing the image I have chosen. A useful edit should let me keep directing the same idea.

Sources: Google Cloud [1] · Google AI for Developers [2]

The light improved. The face was less certain.

Fuser compared five tasks on 6 October: one generation per model, no retries. In its lighting task, 2.1 followed the requested direction; Pro reversed it. In a separate character transfer, Pro preserved the likeness more closely; 2.1 changed hair, glasses and freckles.

This small comparison comes from a commercial creative platform. We have not reproduced it. It cannot establish a universal winner.

A face carries information at several scales. The silhouette may match while the glasses become finer, the expression relaxes or the hair acquires a different rhythm. Viewed alone, the new image can look convincing. Placed beside the approved reference, it can feel like a different casting choice.

That distinction matters when an image must belong to a sequence. I would check whether the character remains recognisable before deciding whether the new lighting is better. Otherwise, I may approve a beautiful shot that no longer belongs to the same scene.

Sources: Fuser [4]

A tiny detail can change the product

In Fuser’s bottle edit, all three models made a smooth cap collar ribbed despite an instruction to preserve it.

For an advertising image, the shape of a cap is part of the object being sold. A plausible replacement still changes the product. The same principle applies to a costume fastening, a distinctive prop or a mark on a set: continuity depends on details that a single attractive frame can make easy to overlook.

I would make those details explicit before editing: these glasses, this label, this closure. Then I would compare them at the size in which the work will actually be used. A social thumbnail and a campaign close-up expose different problems. The approval should belong to the intended use.

Sources: Fuser [4]

Cheaper images. A more complicated bill.

The Standard Gemini API price for 1K image output falls from $0.067 with Nano Banana 2 to $0.0336 with 2.1. At 2K, it falls from $0.101 to $0.0504. The 4K change is smaller: $0.151 to $0.113. Input and text/thinking tokens cost more with 2.1 and are billed separately. These figures describe API image output, not the total request or Flow credits.

Lower output prices make it easier to compare alternatives. A production budget also includes the time spent checking them. If a revision makes me repair the reference, rebuild the layout and inspect every label again, its useful cost has grown beyond the generation charge.

The improvement I care about is fewer broken decisions per revision. That would make cheaper generations useful throughout a project, including the uncomfortable final changes just before delivery.

Sources: Google AI for Developers [5]

Bring the image you are afraid to lose

A useful first trial would start with an image you have already approved. Ask for one clearly defined change. Name the features that must remain. Keep the original beside the result while you judge it. This is a suggested exercise, not a test we have conducted for this article.

Try the lighting note first. Then a different background. Finally, ask for the same character in another composition. Those requests probe different kinds of control. Record the attempts you reject as well as the one you keep; a single successful image tells you little about the effort needed to repeat it.

Nano Banana 2.1 gives us a good reason to revisit this workflow. Its announced gains address practical frustrations, and the early examples reveal specific progress. They also leave the most interesting question open: can the model change exactly what you asked, while protecting the choices that made the image yours? That is the revision worth watching.