Ad
Ad

What inpainting is in AI, and how img2img differs

Two words that get used interchangeably and mean different things. One rebuilds your whole picture; the other is supposed to leave most of it alone.

The short answer

Inpainting in AI changes one named region of your picture and leaves the rest alone. Img2img takes your picture as a starting point and generates a new one that resembles it, with every pixel free to change. Inpainting takes your picture plus a mask and is supposed to redraw only inside the mask. The distinction matters because it decides whether the parts you liked survive: with img2img they may not, and no wording in the instruction changes that.

MethodA comparison of the two operations as they are defined and as our editor implements them. Behaviour claims about our own tool come from the six edits in our [Qwen Image Edit run](/blog/qwen-image-edit/) on 2026-08-11.
img2imginpainting
What you give itAn image and an instructionAn image, an instruction, and a mask marking the region
What may changeEverything in the frameIn principle only the masked region
Typical useRestyle, reimagine, change the whole moodRemove an object, fix a hand, replace a sky
What it costs youFine texture and exact framing, every timeA soft edge where the mask meets the original
What our editor doesThis one. Upload, describe the change, get a re-rendered frameNot offered. There is no mask tool
Free alternatives with a real maskn/aGIMP and Photopea for manual work

The difference in one sentence each

Img2img hands your image to the model as a starting point rather than starting from noise. The model then generates a new image, guided by both your picture and your words. How closely it follows the original is a dial: turn it one way and you get your photo back, turn it the other and you get something unrecognisable.

Inpainting hands the model your image plus a mask, and asks it to fill in the masked area so that it fits the rest. Everything outside the mask is supposed to be copied through untouched.

The words get used interchangeably in product interfaces, which is why people are surprised when a request to change a sweater comes back with a different background.

What "the whole frame is in play" means in practice

We ran six edits through our editor and looked at what moved that we had not asked about. Every single one had something.

Asking for a pair of glasses returned the glasses and a coarser knit on a sweater we never mentioned. Asking to change one word on a sign returned the word and a different wood grain, a slightly larger sign and a softer picture overall. The full set of before-and-after pairs is worth a look, because seeing it is more convincing than reading it.

This is not a defect to be fixed with better prompting. It is what img2img is. Naming what should stay the same does help, measurably, and it is still pressure rather than a guarantee.

Ad

Which one you want

If the answer to "what should stay exactly as it is" is "most of it", you want inpainting with a real mask, and you want a tool that offers one.

If you are changing the mood, the style, the season or the setting, img2img is the right operation and the incidental changes are usually improvements. Restyling is the case where re-rendering everything is the point rather than the cost.

Our editor is img2img. That is a deliberate limit rather than a gap: a mask tool is a different product with a canvas, brushes and undo, and there are good free ones that already do it. For anything that needs pixel-level preservation, GIMP and Photopea will serve you better than we will.

The vocabulary you will meet

A few more terms show up in the same conversations and are worth having straight.

Denoising strength or image weight is the dial mentioned above: how far the result may drift from your input. Low values return something close to your photo, high values treat it as a loose suggestion.

Outpainting is inpainting turned outwards: the mask is the empty space around your image and the model extends the scene.

ControlNet and its relatives constrain the output to a structure taken from your input, a pose, a depth map, an edge map, so the composition survives while the content changes. It is the closest thing to "keep the layout, change everything else".

None of these are exposed in our editor, which takes an upload and a sentence. If you want the dials, a local install is the honest answer, and that is a much larger undertaking than a browser tab.

FAQ

Frequently asked questions

What is inpainting in AI image generation?

Redrawing one region of an image while leaving the rest alone. You supply a mask marking the region, and the model fills it so it fits its surroundings. It is the operation you want for removing an object or fixing a small area of a picture you otherwise like.

What is img2img?

Generating a new image using your existing one as a starting point rather than starting from noise. The result resembles your input to a degree you control, but every pixel is free to change, so fine texture and exact framing usually shift.

What is the difference between img2img and inpainting?

Scope. Img2img may change anything in the frame; inpainting is confined to a mask you draw. If most of your picture needs to survive untouched, inpainting is the operation you need, and no phrasing of an img2img instruction substitutes for it.

Does this site do inpainting?

No. Our editor is img2img: you upload an image, describe the change in a sentence, and get a re-rendered frame back. There is no mask tool. For masked edits, GIMP and Photopea are both free and both do it properly.

Why did other parts of my image change?

Because img2img re-renders the whole frame. In our own six-edit run, every result contained a change we had not asked for, including a sweater texture that shifted when we only asked for glasses. Saying what should stay the same reduces it without eliminating it.

References

Claims last checked .

Ad

Keep reading

More guides

All guides →