AI inpainting, done by naming the part you want changed

Inpainting is an old word from art restoration, where conservators paint in the missing or damaged parts of a canvas so the picture reads as whole again. In digital imaging it came to mean filling a chosen region of a photo with new content that matches its surroundings: repairing a scratch, covering a blemish, replacing a patch of sky. AI inpainting extends the idea from repair to change. Instead of only restoring what was there, you can ask for something new in that region, such as a red raincoat instead of a gray one, or an open door instead of a closed one, while the rest of the picture carries on around it. Most inpainting tools ask you to mark the region with a brush or a mask first and then type what goes inside it. ImageQuarry has no mask step. You identify the part by naming it in your request, which is quicker to start but less exact about boundaries.

Name the part the way a person would. Without a brush, your words are the selection. Use the plain name of the thing and one distinguishing detail: "the blue mug on the left", "the sky above the rooftops", "her scarf, not her coat". If a scene has two of something, give position, color or size so there is only one possible reading.

Describe what replaces it, fully. Inpainting fails quietly when the request says what to remove but not what to put there. "Change the shirt" leaves everything open; "change the shirt to a white linen shirt with rolled sleeves" gives the new content a material, a color and a shape. Say how it should catch the existing light if the change is large.

Draw the boundary with a keep list. A mask stops a change at its edge. Words do that job less exactly, so state the neighbors that must stay: the face next to the collar, the trees below the sky, the frame around the poster. Then look at those neighbors in the new version, because a part change can still affect them.

Expect the edges to need a look. The seams between the changed part and the rest are where problems hide: a collar that blends into the neck, a horizon that wobbles, a frame whose corners have moved. If a seam looks wrong, ask about that seam specifically on the version you prefer.

Change one part per request when it matters. Several part changes in one sentence can work, but when one of them goes wrong you cannot tell which wording caused it. For important images, change one part, review, then move on, and branch from an earlier version if a path goes nowhere.

Use a masking tool when you need exact edges. If a change must stop at a precise line, such as one panel of a garment or a single tile in a floor, a layer editor with a mask gives you that control. Requests in words are best when the part has a clear name and the edge can follow the scene's own boundaries.