All features

Inpainting

AI inpainting on iPhone

Paint over part of an image, describe what should be there, and Onhand regenerates just that area on your iPhone. Masked inpainting runs on the free fast model, on the device, with the source photo staying on the phone.

By Onhand · Updated September 30, 2026

The short answer

Onhand by ArkalEdge is an iPhone and iPad app that lets you paint over part of a photo and regenerate only that area, running fully on the device and offline once the model pack is downloaded. Masked inpainting uses the free fast model, so it needs no subscription; Pro is an optional one-time purchase for more packs. Onhand is pre-launch and coming soon to iPhone and iPad, so join the launch list below to hear first.

One email at launch. No newsletter.

What you get

  • Paint a mask over any part of an image with your finger and regenerate only that area
  • Works on photos from your library and on images you generated in the app
  • Keeps the unmasked pixels exactly as they were
  • Runs on the free fast model pack, so masked editing costs nothing beyond the free download
  • Runs on the device GPU and works offline once the pack is downloaded
  • Every pass is saved to the on-device gallery, and near misses can go back through the Refine loop

What does inpainting do?

Inpainting changes one part of an image and leaves the rest alone. You paint over the area you want redone, describe what should be there, and the model redraws only the painted region. Everything outside the paint stays as it was.

Use it when most of the picture is already right and one patch is not. Typical jobs are replacing an object on a table, redoing a hand or a face that came out wrong, changing the color or cut of a shirt, and turning a flat grey sky into a sunset. It also covers what people usually call removing something: you paint over the object and describe the wall, the grass or the sky that should take its place.

Inpainting is the wrong tool when you want the whole frame to change. Restyling a photo end to end is image-to-image, which works on every model pack in Onhand and has a strength control for how far the result travels from the source. Think of inpainting as a local repair and image-to-image as a new version of the same picture.

How does inpainting work in Onhand?

Onhand gives you a brush. You open an image, paint the area you want regenerated, write a prompt for that area, and generate. The model reworks the masked pixels and keeps the rest of the image.

  1. Pick a photo from your library, or an image you generated in the app.
  2. Paint over the region you want changed, adjusting the brush size for edges and small corners.
  3. Write a prompt that describes the finished region, including its material and its light.
  4. Generate. The pass runs on the device GPU while the app is open.
  5. Keep the version you like. Each version is saved in the on-device gallery with its history.

Masked inpainting runs on the free fast model, a Stable Diffusion-class pack of roughly 2 GB that is free to use after the free app download. It is the only pack in the catalog that edits through a painted mask.

That split is worth knowing before you choose a pack for a job. When the edit has to leave the rest of the frame untouched, the free pack is the one that does it. When you want the prompt following of a Pro pack, the route is to rework the whole image at a low strength and iterate with Refine, accepting that the untouched areas will shift a little on each pass.

How do I get a clean result?

Most disappointing inpaints come from a mask that is too tight or a prompt that names only the new object. Give the model room, and give it the whole region in words.

  • Paint generously. Cover the object and a margin of the surface around it so the model can rebuild the join and the shadow. A mask that hugs an outline tends to leave a halo of the old shape.
  • Describe the finished region, not the edit. "A worn wooden table top in soft window light" works better than "remove the cup".
  • Name the material and the light. The masked patch has to match the rest of the frame, and your words are the only place that information comes from.
  • Reuse the style words from the original prompt when you generated the image in the app, so the patch is made in the same look.
  • Change one thing at a time. Two narrow passes are easier to steer than one mask covering half the picture.
  • Iterate with Refine. Send a near miss back through with an adjusted prompt instead of starting the image over.

Refine keeps every version, so trying a harder prompt costs you nothing you already have. When three passes have not fixed a region, widening the mask is usually more effective than rewording the prompt again, because the model then has enough surrounding area to rebuild the transition.

One thing that does not help: negative prompts. The fast model runs without classifier-free guidance, so a list of things to avoid has no effect on the result. Put the direction you want in the positive prompt instead.

What can inpainting not do here?

Several things people expect from a desktop inpainting workflow are absent, and they are not hidden behind a purchase. They are simply not in the app.

  • No outpainting. The canvas cannot be extended beyond the original frame, so there is no way to widen a crop or add sky above the top edge.
  • No one-tap erase. Nothing removes an object without a prompt. You paint the area and describe the replacement.
  • Text inside images is unreliable. An inpainted sign, label or logo may come back misspelled, and that is a property of the models rather than a setting.
  • No ControlNet, no LoRA, no custom model import. The mask, the prompt and the style presets are the whole set of controls.
  • No AI upscaling. Resize is a plain resize, so an inpainted patch will not gain detail by making the image bigger.

There is also no cloud fallback. If a pack will not run on your device, the app does not quietly send the job somewhere else. Which packs a device can run depends on available memory rather than chip speed, and the developer's test device and floor is an iPhone 14 Pro with 6 GB, where the free inpainting pack runs.

Where does the editing happen?

On the device. The source photo, the mask you paint and the prompt you write are processed by a model running on your iPhone or iPad GPU. Prompts, photos and results stay on the device and are not sent to a generation server.

Once the free pack has downloaded, inpainting works in airplane mode. There is no account to create and no sign-in step. The app does use limited anonymous analytics and Apple and RevenueCat purchase services, which are described in full on the privacy page.

Onhand is a safe-for-work app, and an on-device check runs on results. Model licenses and credits are listed on the licenses page. Onhand runs on iPhone and iPad with iOS 26 or iPadOS 26 and later, and there is no Mac, Android or web version.

What it does not do

  • No outpainting: the canvas cannot be extended past the edges of the original image
  • No one-tap erase: you paint the area and describe what should replace it
  • Text and lettering inside an image are unreliable on these models, so an inpainted sign or label may not read correctly
  • No ControlNet, no LoRA and no custom model import: the mask and the prompt are the controls
  • No AI upscaling, and no cloud fallback if a pack will not run on your device

Frequently asked questions

What is inpainting in an AI photo app?

Inpainting regenerates a selected part of an image and keeps everything else. You mark the area with a mask, write a prompt for what should be there, and the model redraws that patch so it matches the rest of the frame. It is the standard way to fix or replace one region without generating the whole picture again.

Can I inpaint a photo on iPhone without an internet connection?

Yes, in Onhand, once the model pack has been downloaded. The editing itself runs on the device GPU, so a masked pass works in airplane mode. Downloading the pack is the only step that needs a connection.

Does inpainting in Onhand cost anything?

No. Masked inpainting runs on the free fast model pack, which is free to use after the free app download. Pro is a one-time purchase that unlocks additional model packs, and those packs edit by reworking the whole image rather than through a mask. There is no subscription, no credits and no limit on how many images you make.

Can I remove an object from a photo with AI on iPhone?

You can replace it. Onhand has no erase control that deletes an object without a prompt, so you paint over the object and describe what should be behind it, such as the wall, the grass or the sky. For a small object on an even surface that reads as removal. For an object sitting in front of complex detail, expect to paint a wider mask and run two or three passes.

Why does my inpainted area look pasted on?

Usually the mask is too tight. Paint past the edge of the object so the model can rebuild the boundary and any shadow, rather than filling in an outline of the old shape. Naming the light and the material in the prompt also helps the new patch match the surrounding frame.

How long does one inpainting pass take on an iPhone?

The free fast model takes about 7 seconds for a 512 pixel image, measured by the developer on an iPhone 14 Pro, with the method described on the about page. A masked pass is the same generation work, so plan for roughly that order of time. Larger images and larger model packs take longer.

Sources

Facts about other products come from their own listings and documentation, last checked on the date shown. Corrections are welcome at the address on the About page.