Edit your image with Instruct: make smart changes using text

Use a reference image and a simple text prompt to make changes — like changing colors, removing objects, or editing the scene.
Instruct lets you edit an existing image using natural language. You can upload one or more reference images — depending on the model — and describe what you want to change. The system will generate a new version based on your instruction.
You can use it to change the color of an outfit, remove an object, adjust lighting, switch the season, or make many other types of edits.
How to use it
Option 1 – Upload an image from your device
Click Reference image in the prompt panel
Upload one or more reference images (depending on the model)
Set Mode = Reference
Choose Method = Instruct
Write your instruction in the prompt field
Choose a model that supports Instruct (see list below)
Click Create Image
Option 2 – Use an image from your project
Hover over an image in your project
Click the Remix button that appears
The image is added to the prompt panel as a reference image
Make sure the model in the prompt bar supports Instruct (see list below) — change it if needed
Write your instruction in the prompt
Click Create Image
Referencing images with @-mentions
When you have one or more reference images attached, you can refer to them directly inside your prompt using @-mentions. Type @ in the prompt field to pick an image from your references.
Example:
Move the model from @Image-1 to the office in @Image-3. Put glasses on the model from @Image-2
Reference images are named automatically (
Image-1,Image-2, and so on). You can rename any of them to something custom if needed.You can swap the image inside a reference without losing the mention — useful when iterating on the same slot with different visuals.

How it works
The uploaded or selected images become the reference for generation. Your text instruction is used to make targeted changes to that image.
Instruct is available for these Base Models:
Nano Banana
Nano Banana 2
Nano Banana Pro
GPT Image
GPT Image 1.5
Flux.1 Kontext
Other models — including remaining Base Models, Artist Styles, Private Styles, and Exactly 1.0 Agentic — do not support Instruct.
Make sure the aspect ratio of your reference image matches the generation frame. Otherwise, the image may be cropped or resized.

Updated
12 May 2026
