Two edits went to the same photograph. One added a small satin bow at each collar. The other replaced all of both people's clothing. Measured against the original, the small edit moved the faces and the wall ΔE 4.47; the total change of clothing moved them ΔE 4.19. Asking for more disturbed less.
What the routes are
The complaint this starts from is a real one: you have a photo whose expressions you like, you want one thing different, and regenerating the whole picture gives you strangers. So the question is where the boundary sits — how much you can ask for before the photo stops being your photo.
The original. A waist-up photograph of a couple standing side by side in front of a plain pale grey wall, both facing the camera… Both are smiling slightly with their heads turned very slightly toward each other. Soft even daylight from the front. No jewellery, no patterns, no text, no logos anywhere in the frame.
Four routes went out from there. One regenerated from text alone, with the bow written into the same description and nothing attached. Two attached the original as a reference and asked for a change — one small, one large. The fourth took the small change through the tools page's AI Inpainting link instead.
Face score is a normalized correlation against the original face, searched over position so a few pixels of drift do not count against it, expressed as a percentage of what the original scores against itself.
Regenerating from text gives you different people
The text-only route scored 39% and 41%. That is not “a bit off”. When I checked the instrument by measuring one person's face against the other person's face in the same photograph, it returned 38% — so 39% is the score for a stranger. Same description, same bow, different couple.
Attach the original and both faces jump to about 93%, with zero positional drift: the heads land on the same pixels. That is the boundary, and it is a single binary decision. Attach the photo or do not.
What it costs the rest of the picture
Read the third column. Asking for a bow cost the untouched parts of the picture 4.47; asking for two complete garments cost them 4.19. Within the noise of single runs those are the same number, and the larger request is not the more expensive one.
Brightness is distance from the original. The faces and hair glow in all three, most visibly in the two middle routes.
The maps say why. There is no region that stays put. Every measured region changed to some degree in these outputs, and the parts you asked to keep came back at about four units of color distance from where they were — close enough to read as the same photograph, far enough to be a repaint rather than a copy. What makes an edit legible as an edit is that it lands around ten times above that floor: 41 and 56 against 4.2 and 4.5.
So the useful way to think about the request is not “how much am I allowed to change”. It is: everything moves a little no matter what, and your change has to be big enough to be distinguishable from that.
The link on the tools page goes somewhere else
In this 22 September 2026 session, the tools-page route I tested opened the image in Design Studio. I treated that result as a separate workflow rather than as evidence for how CapCut AI Inpainting always behaves.
The same bow instruction produced a more localized-looking result through this route. The untouched areas measured ΔE 2.04, less than half the roughly four-unit change seen in the two reference-edit outputs above, and the left face reached 97% of its original-match ceiling. The difference map was also much darker outside the bows, which means less of the surrounding image changed in this run.
That tighter result came with different output and cost behavior. The Design Studio generation reported 14 credits consumed, compared with the 1-credit amount shown for the comparable reference-edit request in Video Studio. It also returned a 1290 × 720 WebP, while the Video Studio outputs in this test were 2560 × 1440 PNGs.
Those differences make this a workflow choice rather than a simple upgrade. If preserving the surrounding image matters more than resolution or the credit cost shown for the current request, the Design Studio result may be useful to compare. If output size and lower generation cost matter more, the Video Studio reference workflow produced the larger files in this session. These figures describe the routes tested here on this date, not fixed pricing, resolution, or behavior across every CapCut image-editing workflow.
What to do with this
Where this stops
One photograph, one run per route. The 4.47 and 4.19 are not far apart and each comes from a single generation, so the honest claim is that the larger edit was not more expensive, not that it is reliably cheaper. The face scores are far enough apart to carry more weight: 39% against 93% is not a margin that noise closes.
The two reference routes also differ in what was asked, not just in size — a bow is an addition and an outfit is a replacement. And the fourth route ran on a different product with different models, so its lower floor is a property of that pipeline, not proof that masked editing in general is gentler.
There is no setting for leave it alone
The thing that is easy to expect from an editor is a dial: touch a little, touch a lot. The measurements say the dial does not exist. Every one of these routes handed back a complete repaint in which the faces sit four units away from where they were, and the only variable under your control is whether the change you wanted is loud enough to be seen against that. Treat the floor as a fixed cost of asking at all, decide once whether the photograph goes in with the prompt, and then ask for the whole change you actually want rather than creeping up on it — because the creeping does not buy the caution it looks like it buys.
Configuration and cost. The original and the three composer routes: Image, Seedream 4.5, 16:9, 2K, Standard, in Video Studio — the panel declared 1 credit each, and the balance fell by exactly 1 on each of the runs where it could be read cleanly. The fourth route ran in Design Studio, reached through the tools page's AI Inpainting link, which reported 14 credits consumed on completion. Six generations in total: five at 1 credit and one at 14, so 19 credits, 2026-09-22. One of those five is a discard, explained below.
The balance could not be used as a running check today. It rose twice during the session without a generation between the readings — 23,665 to 23,714, and 23,714 to 23,863 — so credits were being added to the account while the work was running. Per-step deltas are reported only where the reading before and after a single send was clean; elsewhere the figures are the declared ones.
One generation was discarded. The whole-outfit request first went out with the wrong reference attached: the canvas had shifted by one card width and the click landed on the text-only version rather than on the original. It cost 1 credit, its result is not in this article, and the run was repeated against the correct source.
Files. Original and the three composer outputs, 2560×1440 PNG. The Design Studio output, 1290×720 WebP. Prompt text is quoted as sent, abbreviated only where marked.
How this was measured. Face identity is normalized cross-correlation on locally contrast-normalized images, searched over a window so that position drift does not register as a difference. That search matters: the fixed-box version of the same measure fell from 1.00 to 0.76 on a two-pixel shift and to 0.27 on six, which would have made drift look like a change of face. The search version returned the same score for shifts of 2, 6, 15 and 30 pixels and recovered each displacement exactly, stayed flat under brightness changes of ±30%, and scored 0.38 when handed a different person's face — which is the number the text-only route matched.
Color distance is CIELAB ΔE, checked against a pair of known distance 10, which returned 10.0. Every sampling box was drawn back onto all three images and inspected before any number came out of it.
One download was caught and thrown away. The first file pulled from the Design Studio canvas was pixel-identical to the source image — every one of 3.7 million pixels matched, so it was the input rather than the output, sitting on the same canvas at a larger size. The measured output is the 1290×720 file.