Nano Banana

Nano Banana is Google’s image generation and editing model, known for instruction-based edits that leave the rest of the image intact.

Last verified 2026-08-26

Nano Banana is Google’s image model, notable less for generating images from scratch than for editing existing ones by instruction — and for how well it leaves everything you did not ask about alone.

Instruction editing

The usual problem with editing an image generatively is collateral damage: ask to change the sky and the subject's face subtly changes too, because the model regenerated the whole frame. Nano Banana's strength is applying a described change while preserving the rest, which makes iterative work practical:

• Change a background without touching the subject.
• Alter clothing, colour or an object in place.
• Remove or add an element while the lighting stays consistent.
• Adjust one region while the rest of the frame is untouched.

This also makes it useful for identity work: because it can edit while preserving facial structure, it fits pipelines that need a specific person to survive an edit.

Writing edit instructions

Say what should change and what must not:

• Name the target precisely — "the red mug on the left", not "the object".
• State the change plainly.
• Name what to preserve when it is at risk — "keep the face and lighting unchanged".
• One change at a time. Stacked instructions in a single pass interfere; sequential edits are more predictable.

Versions

Several Nano Banana variants exist, including a Pro tier and a second generation. They differ in quality, speed and cost, and Higgsfield surfaces more than one. As with every model here, confirm which variant a given surface calls rather than assuming the newest.