Make controlled edits to AI images

The image editor lets you mark up an AI-generated image with masks, references, and arrows, then feed the marked-up version to a model with a short prompt. The mask says where. The prompt says what. Branching Create nodes lets you try different edits without losing the original. Time: 10 minutes You’ll need: A generated image to edit (or any image on your canvas)

​

1. Open the image editor

Double-click any image on your canvas to open the editor. The main canvas fades, a tool tray appears at the top, and a layers panel slides in on the right. The original image sits locked at the bottom layer so you can’t accidentally paint over it. Each stroke or pasted image becomes its own layer.

​

Tools

• Mask (M): paint a region for the model to edit. Default tool.
• Brush (B): regular pen. Set size and colour from the tray.
• Eraser: works on whichever layer you have selected.
• Arrow (A): draw a line with an arrowhead. Good for pointing.
• Text: add text on the canvas.
• Crop: drag a rectangle, optionally lock to a fixed ratio.

For the full image editor reference see Create node, Image Editing.

​

2. Mask the area you want to change

Press M to make sure the mask brush is active, then paint over the area of the image you want the model to edit. The mask layer is always red. Hit save when you’re done. The image node updates and you can flip back to your canvas.

​

3. Branch a Create node and write the prompt

Drag from the masked image’s output to a new Create node. Pick Runchat Image for fast iteration, or Nano Banana Pro for ship-quality work. The prompt should have three parts:

1. Name the mask. Start with “In the red mask area…” to anchor the model to the region you painted.
2. Be concrete. “A thoughtfully designed outdoor space with colourful furniture” beats “make it nicer”. Vague verbs like “improve” or “fix” almost never work.
3. Point at a reference. When you have a reference image, mention it: “…using the furniture from the reference image”.

One idea per prompt. Read it back as a single instruction you could give a retoucher.

​

4. Pasting a reference into the image

When the model struggles with scale between a separate reference and your target, paste the reference into the same image as a new layer:

1. Copy the reference (or drag it onto the canvas first, then copy from there)
2. Double-click into the target image to open the editor
3. Paste, it lands as a new layer
4. Tuck it into a corner where it covers something you don’t mind losing
5. Save

Then prompt: “Use the bottom-left as the reference.” Same coordinate space, better scale.

​

5. Try one edit at a time

Single edits beat compound ones. Asking for three things at once usually gets you one right and two wrong. Branch a new Create node for each change and layer them. Simple unrelated edits, adding a person plus interior lighting, can compound fine. Four things at once cannot. When a result is close but not right:

• Adjust the mask and rerun. The mask is the cheapest variable to change.
• Switch to a stronger model like Nano Banana Pro on the same prompt.
• Mask the leftover artefact and run a cleanup edit on top.

Every branch is non-destructive, your original is always upstream.

​

Quick tips

• The mask brush is the default tool. You don’t need to switch.
• Smaller brush for fine edits like adding a figure.
• Pan and scroll the editor viewport like the main canvas.
• Backspace deletes the active layer.
• Layer history is preserved across sessions, your mask stays editable when you reopen.

​

Next steps

• Render from a Rhino screenshot, generate the baseline images these edits build on
• Train a custom LoRA, for systematic style consistency across edits