Best AI inpainting tools in 2026: selected-area edits

AI inpainting used to mean one thing: paint a mask over the part of the picture you dislike, type a short prompt, and let the model redraw inside the mask. In 2026 that's only one of several routes to the same result. Some tools still hand you a brush. Others let you name the region in plain language — "the second bottle from the left", "the sign above the door" — and edit only that area while the rest of the frame stays byte-for-byte familiar. Both approaches count as inpainting if what you get back is a selected-area edit rather than a fresh image.

This roundup covers eight ways to do selected-area editing in 2026, starting with the multi-model route on OmniArt. For each one: the kind of region control it gives you, the job it actually suits, and where it runs out of road. No tool here wins every brief, and the honest limits matter more than the feature lists.

What counts as inpainting now

Three jobs hide under the same word, and mixing them up is the most common reason an edit disappoints.

Removal is a reconstruction problem and is often better served by a small, specialized model. Replacement and local adjustment are generation problems and want a capable image model with tight regional obedience. Choose the tool for the job, not for the brand.

1. OmniArt — instruction-and-region editing across several models

OmniArt's route to inpainting-style results is instruction-and-region editing inside the image models themselves rather than a mask-brush canvas. You describe the region and the change in the prompt, and the model edits that region while leaving the rest of the frame untouched. Three models on the platform carry most of this work.

Seedream 5.0 Pro is the precision option. It responds to coordinate and positional editing — a box, a point, an arrow, or a literal coordinate naming the target area — plus sketch editing, where a rough shape you supply becomes the spatial reference for a described object. Anchor editing pins a small or ambiguous target so the model doesn't wander to a neighboring object in a busy frame. It also reads exact hex codes and named materials, does multi-image fusion across references, and can split an output into a background layer plus separate transparent element layers.

GPT Image 2 offers a natural-language editing surface: you refine an existing image by describing the change in sentences, which suits edits that are easier to say than to outline. Nano Banana 2 handles reference-guided edits, where another image supplies the look, object, or material you want carried into the target region.

• Selected-area control: described regions, coordinates, anchors, sketch references, hex-exact recolors
• Best for: brand-accurate product edits, multi-reference composites, layered assets you'll recompose downstream
• Honest limit: there is no brush or mask UI on OmniArt. Region control lives in prompt grammar, so precision depends on how clearly you can name the area — and some model-side inputs like sketch and anchor placement remain upstream API workflows rather than OmniArt canvas controls

2. Adobe Photoshop generative fill — the mask-brush standard

Photoshop remains the reference implementation of classic inpainting. Lasso or brush a selection, type a short prompt, get three variations on their own generative layer, and keep iterating non-destructively. Because the selection is pixel-exact, it handles the cases that language struggles with: an irregular shape, a gap between two overlapping objects, a sliver along an edge.

• Selected-area control: pixel-exact masks, feathering, per-variation layers
• Best for: retouching, extending canvases, compositing where the mask must be exact
• Honest limit: subscription-bound and desktop-first; generated fills can read softer than the surrounding photography, so large replacements often need a second pass

3. Stable Diffusion inpainting in ComfyUI or A1111 — maximum control, local

The open-source stack gives you knobs nobody else exposes: denoise strength, mask blur, inpaint-at-full-resolution, dedicated inpainting checkpoints, and ControlNet conditioning to hold pose, depth, or edges inside the masked area. It runs on your own hardware, which matters when the source images can't leave your machine.

• Selected-area control: full mask parameters plus structural conditioning
• Best for: repeatable production pipelines, sensitive material, style-specific fine-tuned models
• Honest limit: the steepest setup and learning curve on this list, and results are only as good as your checkpoint choice — expect real time spent before the first usable edit

4. Midjourney region editing — redrawing inside a stylized frame

Midjourney's region tools let you select an area of a finished generation and re-prompt just that part, which keeps its distinctive rendering intact instead of forcing you to export into a separate editor with a different aesthetic. For illustration and concept work, staying inside one visual system is worth more than mask precision.

• Selected-area control: rectangular and freehand region selection with a re-prompt
• Best for: concept art, stylized keyframes, iterating on a look you already like
• Honest limit: weaker on photographic realism and text; region edits inherit the model's stylistic pull, so literal, documentary-accurate replacements are hard to force

5. LaMa-class object removers — cleanup without invention

Tools built on inpainting-specific architectures — Cleanup.pictures and the various open-source LaMa front-ends among them — do one job: brush over an object and get plausible background back. They aren't trying to generate a new subject, which is exactly why they're good at removal. Power lines, tourists, stray cables, and logo plates disappear without a hallucinated replacement sneaking in.

• Selected-area control: brush mask, removal only
• Best for: high-volume cleanup, photo prep before a generative edit
• Honest limit: no replacement or local adjustment at all, and large masks over complex texture (crowds, dense foliage, patterned tile) smear

6. Canva Magic Edit and Magic Eraser — inpainting inside a design file

Canva's value is location, not capability. The edit happens in the same file as the layout, the type, and the brand kit, so a social post can be fixed without a round trip through a separate editor and a re-import. Non-designers on a team can do a passable object swap without learning masking.

• Selected-area control: coarse brush selection, short text prompt
• Best for: social and marketing assets, teams with mixed skill levels, fast turnarounds
• Honest limit: the coarsest region control here; fine edges, small objects, and print-resolution output are outside its comfortable range

7. Photoroom — product and ecommerce edits at volume

Photoroom is tuned for the commerce case: cut the product, clean the surface, remove a reflection or a price sticker, replace the background, and batch the whole thing across a catalog. Its object-aware selection usually finds the product without much brushing, which is what makes hundreds of SKUs tractable.

• Selected-area control: automatic object detection plus manual refinement
• Best for: marketplace listings, catalog cleanup, repeatable batch edits
• Honest limit: narrow by design — outside product photography its selection heuristics have much less to work with, and template-driven output can look uniform across a catalog

8. Google Photos Magic Editor — phone-first, on the roll

For edits that need to happen where the photo already lives, Magic Editor handles removal, repositioning, and background changes with a tap-to-select model instead of a brush. It's the tool most people will actually use, because it's already installed and the source file never gets exported anywhere.

• Selected-area control: tap and lasso object selection, guided suggestions
• Best for: personal photos, quick social fixes, edits on a phone
• Honest limit: limited prompt control, no layer output, and resolution and fidelity ceilings that make it unsuitable for client or print delivery

Picking by job

Prompt patterns that travel across tools

The instruction habits that lift results are broadly portable, whether you're typing into a mask dialog or naming a region in a sentence.

1. Name the target the way a person would say it out loud. "The second bottle from the left", "the sign above the door", "row 2, column 3" all beat abstract percentages.
2. Say what changes and what doesn't. One clause for the edit, one clause protecting the rest of the frame.
3. Be numerically specific where numbers exist. Hex codes for color, named materials for surfaces — "a bit warmer" is not an instruction.
4. Describe light and contact, not just the object. A replacement lands when you state the key light direction and where the object touches the surface below it.
5. Split compound edits into separate passes. Three changes in one instruction is how a local edit turns into a whole-frame regeneration.

What this list leaves out

Single-purpose mobile apps that wrap a public removal model, and enterprise DAM tools where inpainting is a checkbox behind a sales call, are both out. So are watermark removers, which are a legal problem more than a technical one. The bar for inclusion was simple: real region control, a job it does better than the alternatives, and availability to an individual creator today.

Getting started on OmniArt

Selected-area editing on OmniArt starts the same way as any other generation: bring the image into the image workspace and write the region and the change into the prompt. For the full pattern library — coordinate edits, sketch and anchor references, hex-exact recolors, multi-image fusion, and layer separation — the Seedream 5.0 Pro prompt guide has copy-pasteable examples for each mode.

If your edits read better as sentences than as coordinates, GPT Image 2 is the one to try first — its natural-language editing surface is built for describing a change rather than outlining it. Having both models in one workspace is the point: when one won't respect a region, the next one is a dropdown away rather than another subscription.