Use the AI tools that run on your device
Remove background, select subject, object select, remove object and expand with AI fill: what each does, which model it uses, the one-time downloads, and what never leaves your browser.
Beginner5 min readUpdated 25 September 2026Editor · On-device AI
On this page
The Editor has five AI tools, and all of them run in your browser. Your images are never uploaded, there are no credits and nothing is charged. This article covers what each tool does, what it downloads the first time, and what to try when the result is not right.
The tools
| Tool | Where | What it does |
|---|---|---|
| Remove background | Properties > Quick actions, or Layer menu | Hides everything but the subject of an image layer with a mask |
| Select subject | Select > Subject, or the options bar of any selection tool | Selects the main subject of the whole visible design |
| Object select | Tool rail, key W | Selects the object inside a box you drag |
| Remove object | Tool rail, Shift+J, or Filter > Remove object (AI, on device) | Fills in whatever you paint over, as if it was never there |
| Expand with AI fill | Image > Expand with AI fill… | Enlarges the canvas and paints the new edges to continue the picture |
Remove background has its own article: Remove a background on your device.
The models and their downloads
Each tool uses a model. A model is downloaded once, the first time you use a tool that needs it, then kept in your browser and used offline from then on.
| Model | Download | Licence | Used for |
|---|---|---|---|
| MODNet | About 26 MB | Apache-2.0 | Remove background and select subject (best on people) |
| BiRefNet lite | About 115 MB | MIT | Remove background and object select (any subject). Needs WebGPU; without it the MODNet model is used instead. |
| LaMa | About 208 MB | Apache-2.0 | Remove object and expand with AI fill |
Before a model downloads for the first time, the browser shows a prompt with its name and size, saying it runs on your device, nothing is uploaded, it is free, and it works offline after the first download. Click OK to go ahead. If you cancel, nothing is downloaded and you are asked again next time. Once you say yes, you are not asked again for that model.
Help > AI on this device lists every model with its size and licence. The same list is under Edit > Preferences… > AI and privacy. Both have Remove downloaded models, which deletes the cached models and resets the prompts, useful if you need the storage back.
Select subject
Select > Subject runs the MODNet model on everything visible and turns the result into a selection. It respects the selection mode in the options bar, so with Add to selection chosen it adds the subject to what you already have. When it finishes it suggests Select and mask for hair and soft edges. The same Select subject button is in the options bar of every selection tool and inside the Select and mask workspace. See Selections.
Object select
- Press W for Object select.
- Drag a box around the object. Leave a little room round it.
- The Editor finds the object's edges inside the box and selects it. Anything the model finds outside the box is left out.
Hold Shift while dragging to add to the selection, Alt to subtract. If your browser has WebGPU, Object select uses the any-subject model (BiRefNet lite); otherwise it uses MODNet, which is best on people. Very small boxes are ignored.
Remove object
- Select the image layer that holds the thing you want gone.
- Press Shift+J (or choose Filter > Remove object (AI, on device)).
- Set the brush size in the options bar and paint over the object, covering it completely, shadow included.
- Let go. The Editor shows Removing and fills the area to match its surroundings.
Remove object works on a square area around what you painted, scaled to 512 by 512 pixels for the model, then puts back only the painted part with a soft edge, so the rest of the photo keeps its full resolution. It works best on objects that take up a modest part of the frame. A large area on a big photo can come back softer than the pixels around it.
If the model cannot start in your browser, the Editor says The AI remover could not start in this browser, so a simpler fill was used. and falls back to the same patch fill that Spot heal uses. If the area is too big even for that, it says That area is too large to fill. Try a smaller area. For small blemishes, Spot heal (J) is quicker and needs no download.
Expand with AI fill
Image > Expand with AI fill… opens the canvas size window at 125% of the current size with Fill the new area with AI ticked. Set the size and anchor, then click Apply. The new area is filled on its own layer, called AI fill, at the bottom of the stack, and the Editor says New area filled on its own layer at the bottom. Hide or paint on it to adjust. You can also tick the same box in Image > Canvas size…. See Crop, resize, rotate and flip.
It uses the same 512 by 512 approach as Remove object, so modest expansions look best. Expanding a photo by a little on each side works better than doubling it in one go.
Speed and browsers
- The first run of each model is the slowest, because it downloads and starts the model. Later runs are quicker.
- Where the browser supports WebGPU, Remove object and expand use it, and fall back to a slower path that works everywhere. The any-subject model needs WebGPU to be usable.
- The models are large for a phone. They can work there, but expect longer waits, and use Wi-Fi for the first download.
- See Browser support for which browsers have WebGPU.
Common problems
- Could not load the model. The first download failed. Check your connection and try again.
- The subject selection includes too much. Select subject looks at the whole visible design. Use Object select and draw a box, or tidy the result in Select > Select and mask….
- The removed area looks smudged. Paint a little beyond the object's edge so no fringe is left, and try removing a big object in two or three smaller passes.