Chitra Studio

Editing17 October 2026

How to change one part of an AI image without touching the rest

Draw a box, say what changes, keep every other pixel. Comparing Nano Banana 2 Edit and GPT Image 1 Mini.

I had an AI generated image of an opera singer on a gondola, and I wanted her gown green instead of red. Nothing else needed to change.

Image models will make that change if you ask, but they often redraw more than you asked for. In the test below, one model also changed her face, her pose and the water around her. If a client has already approved the picture, or it has to match others in a set, that means starting over.

The fix is to tell the model where to work, and then make sure nothing else can change.

Mark the area, then say what changes

Editing a marked area in Chitra Studio. While the model works, the box glows and a timer shows which model is editing.
  1. Open the picture and choose Edit.
  2. Drag a box around what should change. Each box gets a number.
  3. Type what should happen in that box, such as "make the gown emerald green". Add more boxes for more changes, each with its own note.
  4. Choose an edit model and press Edit.

What happens behind the box

The box is not just a hint to the model. Three things happen when you press Edit:

  1. Chitra cuts out a piece of the picture around your boxes, in a shape the model accepts.
  2. It draws your boxes on a copy of that piece, in red with their numbers, and adds to your note: change only what is inside the boxes, and do not draw the boxes in the result.
  3. When the result comes back, it is laid over your original only inside the boxes, with a soft edge that stays within each box.
The opera singer picture with a red box numbered 1 drawn around her gown
The box as it was drawn for the test below. The model works from a marked copy; your original file is never changed.

The third step is the one that matters. However the model behaves, nothing outside the box can change, because the rest of the picture is never taken from the model's result at all.

The test: same box, same words, two models

The picture is a Nano Banana 2 image of an opera singer on a gondola on the Grand Canal. I boxed her gown, typed "make the gown emerald green" and ran it on two edit models. In the maps below, everything the model changed is pink.

Two results side by side. Nano Banana 2 Edit turned only the gown green. GPT Image 1 Mini turned the gown green but also redrew the gondolier, the water and the bridge inside the box. Below each, a grey copy of the picture with the changed pixels in pink.
Top: the results. Bottom: what changed, in pink. Grey means identical to the original.
Close-ups of the singer's face: the original, the Nano Banana 2 Edit version, which is the same, and the GPT Image 1 Mini version, with a different face and pose
Inside the box, up close. Nano Banana 2 Edit kept her; GPT Image 1 Mini redrew her face and pose.

Nano Banana 2 Edit took about 20 seconds and cost 12 credits ($0.12). GPT Image 1 Mini took about 45 seconds. I ran Nano Banana 2 Edit three times with slightly different boxes, and all three changed only the gown.

Tips for clean edits

Try it on your own pictures

Chitra Studio is a Mac app for making and organising AI images, video and sound with more than 60 models. Marking areas to edit is built in, alongside Animate, Add Sound, voices and music.

Download for Mac14-day free trial · 50 free credits · no card