Edit Part of an AI Image, or the Whole Frame?

Updated 2026-10-27·3 min read·Rendering

Why region editing disappoints

Selecting a region feels precise. In practice, the tools often don't edit the region you drew. In our tests the editor padded a selection hard: a box about 107 pixels wide became about 289. In a frame full of similar subjects, that padding reaches a neighbor, and the model redraws or duplicates whoever it swallowed.

We saw it while removing one worker from a dock full of eight:

• A tight box around him got widened to include the next man. He came back with a smeared head and a ghosted copy, plus some invented boxes.
• A larger box produced a headless figure inside the edited area.
• A plain sentence describing the change, applied to the whole frame with no region, came out clean. There was no seam and no ghost.

One catch: the whole-frame edit removed two figures when we asked for one. The frame stayed coherent, which the region attempts never did, but always count after an edit.

Crowded frames are the worst case for region editing, and they're also when you're most likely to reach for it.

Masks aren't a shortcut

Mask inputs look like the answer. We found one model silently ignored the mask: it deleted a car and tripled the size of the subject while appearing to work. Another honored the mask but re-rendered and resized the whole frame anyway. Test a model's mask handling before trusting it.

When a region really is the answer

Size is the dividing line. On a wide drone shot, we needed to turn one traffic signal that took up about 1% of the frame. A written description didn't do it, and neither did a box drawn on the image or different wording. What worked the first time was to crop the area, enlarge it about six times, edit the crop, and composite it back. The model couldn't see something that small until it was big.

How to choose

1. If you can name the change in a sentence, such as "Remove the man in the red jacket" or "Turn her to face the camera," edit the whole frame with that sentence.
2. Check the result. Count people, check hands and faces, look for duplicates.
3. If the subject is tiny, roughly 1% of the frame, crop it, enlarge it, edit it and paste it back.
4. Save as PNG if your claim is "nothing else changed." A JPEG re-save re-encodes the whole image and shifts pixels you didn't touch.
5. If you send several images to an edit, describe them in the right order. Telling a model the first image is the full scene when it's actually a crop leads it to copy the wrong one.

An example

A storyboard frame shows a café with an extra waiter who shouldn't be there. Don't mask him. Write "remove the waiter standing behind the counter," edit the whole frame, then check that nobody else vanished. If the problem were a misspelled sign far in the background, crop the sign, enlarge it, fix it, and paste it back.

In ShowMaker

Shoot with AI works the whole-frame way: describe in plain words what's wrong with a shot, and it redraws the frame and writes the motion prompt, showing the credit cost before you spend. Every shot is checked for things like a frame showing something the text never mentioned. Read more on the storyboard, or see how we keep characters consistent through edits.