Nano Banana Image Editing Prompts: Add, Remove, Replace, and Restyle

Nano Banana is Google’s name for Gemini’s native image generation family. Google’s official documentation describes conversational image generation and editing from text, images, or both, including adding, removing, or modifying elements, changing style, adjusting color, and working with multiple references. Model versions differ in speed, resolution, reference handling, and precision, so check the current model enabled in your workspace.

The editing principle is consistent: identify the reference, define the smallest change, protect everything else, and review the output against a visible checklist.

Write edits as deltas

A delta describes the difference between the input and desired output. Instead of rewriting the full scene, say: “Keep the person, pose, crop, lighting, background, and camera unchanged. Replace only the beige jacket with a dark green wool coat of the same length and fit.”

The shorter the delta, the easier it is to tell whether the edit succeeded.

Add an object

“Add one small ceramic cup on the empty desk area to the right of the notebook. Match the existing overhead perspective, soft morning light, warm color balance, realistic scale, and contact shadow. Preserve the hands, laptop, notebook, table texture, and every existing object.”

Specify location, scale, light, and contact. Without those constraints, an added object may float or cover approved content.

Remove an object

“Remove only the floor lamp behind the sofa. Reconstruct the wall, baseboard, and soft window shadow that would naturally continue behind it. Preserve the sofa, rug, person, crop, color, and all other furniture.”

Describe what exists behind the removed object so the reconstructed area belongs to the scene.

Replace an object

“Replace the red plastic chair with a light-oak dining chair of the same size and position. Match the original camera perspective and contact shadow. Keep the table, room, people, lighting, and composition unchanged.”

Protect size and position when the replacement must not alter the layout.

Change one material or color

“Keep the product geometry, label, cap, camera angle, background, and lighting unchanged. Change only the outer case material from glossy white plastic to brushed silver aluminum. Add fine directional metal grain and one broad softbox reflection while preserving every edge and opening.”

A material edit needs new light behavior, not just a color word.

Restyle an image while preserving content

“Transform the supplied city photograph into a paper-cut editorial illustration. Preserve the road layout, building positions, number and location of people, vehicle direction, and crop. Use layered paper shapes, subtle fiber texture, shallow cast shadows, and a limited blue, cream, and orange palette. Do not add signs or new objects.”

Content constraints and style rules are separate. Repeat both when a multi-turn edit begins to drift.

Combine multiple references by role

When the enabled Nano Banana model accepts multiple references, tell it what each image contributes. Example: “Use reference 1 for the person’s facial identity and hairstyle. Use reference 2 for the navy coat design and fabric. Use reference 3 for the train-platform environment and lighting. Create a medium-full portrait that combines only those roles; do not copy the face or body from references 2 or 3.”

Number the references and assign one role to each. “Combine these images” leaves identity and style ownership ambiguous.

Create a campaign variation

“Preserve the approved product, label, camera height, crop, and shadow from the reference. Change only the background from pale beige to a dark botanical greenhouse at dawn. Add one blurred leaf in the far foreground for depth, keep the label unobstructed, and leave clean upper-right space for copy.”

This prompt protects the commerce asset while changing campaign context.

Run sequential edits from an approved state

Save each accepted result. Edit background first, approve it, then adjust color, approve it, then add one prop. If the third edit drifts, return to the second approved image rather than continuing from a flawed generation. Multi-turn convenience does not remove the need for version control.

Protect identity across edits

Maintain a compact identity block: face shape, eye spacing, hair silhouette, age, body proportions, and one signature detail. For major pose or camera changes, add another approved reference angle. Do not rely on “same person” when the new view reveals anatomy absent from the source.

Review provenance and usage rights

Google states that generated Nano Banana images include SynthID watermarking. You are still responsible for having the rights to uploaded references and for reviewing whether an edit could mislead, infringe, or misuse a person’s identity or protected brand material.

Use a five-question edit check

Did only the requested region change? Are identity and geometry stable? Do new objects match perspective and light? Did any text or logo drift? Can the result support the intended crop and use? If the answer to one question is no, revise that exact requirement.

Nano Banana editing works best as a controlled sequence of small deltas. Assign reference roles, preserve approved details, change one production variable, and continue only from an output that passes review.