Nano Banana Pro is best understood as an image editing workflow, not just a text-to-image model. You can start with a photo, illustration, product shot or design and ask for a controlled change. The quality of the result depends heavily on how clearly you separate the requested edit from the parts that must remain stable.
This guide uses the Imagine AI image generator and AI image tools as the workflow entry points. The exact model name and available resolution options depend on the models enabled in your account.
When to use Nano Banana Pro
Nano Banana Pro is a useful choice when your starting point already contains most of the visual information you need. Common tasks include:
- changing a background or time of day;
- restyling a room without moving walls or windows;
- changing clothing while preserving the person;
- cleaning up an old photo;
- transferring a palette, material or lighting direction;
- updating a poster or product image while keeping the layout.
If you are creating a completely new scene with no reference, start with a normal text-to-image prompt. If you need to edit one local region, use a mask or a dedicated tool where available.
The “change only” pattern
The most reusable editing pattern is:
Change only X. Keep A, B and C unchanged. Do not add D.
For example:
Change only the wall color from pale blue to warm white. Keep the room layout, windows, furniture, camera angle, shadows and natural light unchanged. Do not add decorations or text.
This is more precise than saying “make the room nicer.” It gives the model a target and a preservation checklist.
When the edit is local, identify its location:
Remove the person in the far-left background. Rebuild the trees and path behind them so the perspective and lighting match the original. Keep the main subject, face, clothing and camera framing unchanged.
How to use reference images
Before uploading references, decide what each image is responsible for. You might use:
- one image for the person or product identity;
- one image for clothing;
- one image for the target room style;
- one image for color, texture or lighting.
Write these roles directly in the prompt:
Reference 1 is the person to preserve. Reference 2 is the clothing design to transfer. Apply the clothing from reference 2 to the person in reference 1. Keep the face, hairstyle, pose, body proportions, background and lighting from reference 1.
Avoid ambiguous prompts such as “combine these images.” They do not tell the model which image is the subject and which image is the source of style or material.
For a person, describe the fixed attributes that matter: facial features, skin tone, hairstyle, body proportions, expression, pose and camera distance. For a product, name the label, silhouette, material and distinctive parts. This helps the model understand what “preserve the subject” means.
Preserving a poster or product layout
For layout-sensitive work, mention the design system, not only the object:
Replace the background photo with a quiet evening street scene. Preserve the poster’s canvas size, headline position, text hierarchy, logo placement, margins and color palette. Keep the exact headline “CITY NIGHTS” once and add no other words.
After generation, zoom in and inspect the small text. AI image editing can preserve the general layout while still changing a letter, spacing or logo. If the copy is exact and commercial, keep the final text check in your production workflow.
For a dedicated text change, use . For a product asset, use .
A controlled editing workflow
1. Start with a clean reference
Use the highest-quality source you have. A clear subject, visible edges and consistent lighting give the model more information than a compressed screenshot.
2. Make one meaningful change
Change the outfit, background or lighting first. Do not also ask for a new camera angle, new pose and a completely different art style unless the larger transformation is intentional.
3. Repeat the preservation rules
On every revision, repeat the attributes that matter. The model sees the previous output as a new input, so do not assume it will remember an earlier constraint perfectly.
4. Inspect the edges
Look at hair, hands, product labels, object boundaries, reflections and shadows. These areas often reveal whether the edit blended naturally.
5. Move to a dedicated tool when the task is narrow
A focused tool gives you a clearer input contract. Use for removing clutter, for damaged photos, or for interior changes.
Prompt examples
Outfit change
Use reference image 1 as the person and reference image 2 as the clothing. Change only the outfit. Preserve the person’s face, hairstyle, skin tone, body proportions, pose, background and lighting. Make the new clothing fit naturally with correct folds and shadows. Do not add accessories or text.
Product background replacement
Keep the uploaded bottle exactly the same, including its silhouette, cap, label, colors and material. Replace the background with a warm neutral studio sweep. Add a soft contact shadow below the bottle and a subtle reflection on the table. No extra bottles, hands, labels or words.
Room restyling
Restyle the room in reference image 1 as a calm Japanese-inspired living room using reference image 2 only for the color palette and material direction. Keep all walls, windows, doors, camera position and room geometry unchanged. Change only the furniture, textiles and decor.
Common failure modes
The model changes the face. Add identity constraints and reduce the number of other changes in the same request.
The product label becomes unreadable. State that the label must remain unchanged, then check it at full size. If the label is critical, use a final design pass.
The background looks pasted in. Describe the light direction, contact shadow, perspective and reflections instead of asking only for a “realistic” background.
The model copies the wrong reference. Give each reference a role and say which image supplies the subject.
The entire image is rewritten. Use a local mask or a dedicated edit tool and say “change only the selected area.”
Start editing in Imagine AI
Open the , upload your reference images and assign each one a role in the prompt. If your task is a standard repair or transformation, start from so the uploader, validation and edit controls match the job.
For additional model context, compare Google’s and its . A useful community pattern is the public X workflow showing GPT Image output being refined with a Nano Banana edit: .
