All features

Image to image

Image to image on the iPhone

Onhand takes a photo you already have and turns it into a new image in the style you describe. A strength control decides how much of the original survives, every model pack can do it, and the work happens on the device.

By Onhand · Updated September 30, 2026

The short answer

Onhand by ArkalEdge is an iPhone and iPad app that redraws a photo from a text prompt, with a strength control for how much of the original survives, and it runs fully on the device and offline once a model pack is downloaded. No subscription is needed: image to image works with the free fast model, and 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

  • Start from any photo in your library and redraw it from a prompt, on every model pack in the catalog
  • A strength control that sets how much of the original picture the result keeps
  • A Refine loop that sends the result you are looking at back through as the new starting image, as many times as you like
  • Every version kept in the on-device gallery, so you can go back to an earlier pass
  • Style presets, plus on-device resize and background removal to prepare a source photo
  • Generation and editing run on the device GPU, so the photo, the prompt and the result are never sent to a generation server
  • Works in airplane mode once a model pack has been downloaded

What is image to image, and what is it good for?

Image to image starts from a picture instead of a blank canvas. You pick a photo, write a prompt for what you want it to become, and the model redraws the whole frame so the result follows your words while holding on to part of the original. In Onhand this works with every model pack in the catalog, including the free one, and it runs on the iPhone itself.

It is the right tool whenever you already have the composition and want a different treatment of it. Text to image is better when you have nothing to start from and want the model to invent the whole scene.

  • Restyle a photo: a portrait as an ink drawing, a street scene as a watercolor, a landscape as a painted poster.
  • Turn a sketch into a scene: photograph a rough pencil drawing and let the model fill in materials, light and detail.
  • Change the time of day or the weather: a midday street at dusk, a dry road under rain, summer trees in snow.
  • Change the setting behind a subject while keeping roughly where everything sits in the frame.
  • Try a look before you commit: run the same photo through several style presets and keep the one that works.

The photo you start from, the prompt you type and the image that comes out stay on the device. There is no account and no upload step for generation or editing. What the app does collect is limited anonymous analytics and the purchase services Apple and RevenueCat provide, which are described on the privacy page.

How does the strength control work?

Strength decides how much of the original photo survives. It is the one dial that matters most in image to image, and it is the difference between a recolor and a completely new picture. Low strength keeps the composition. High strength reinvents the scene.

  • At the low end, the model barely moves. Faces, poses, edges and layout stay where they are, and you get a change of palette, texture or grain. Use this when the photo is already right and you only want a finish on top of it.
  • Through the middle, shapes stay recognizable while surfaces, lighting and small details are rewritten. A person is still that person in that pose, wearing something the prompt described. This is the range most restyling work lives in.
  • At the high end, the original acts as a rough sketch. The model keeps the broad arrangement of masses and color and invents almost everything else. Use this for sketch-to-scene work, or when the prompt matters more than the photo. Push it to the top and the source stops mattering much at all.

The practical method is to start in the middle, look at what came back, and move one step at a time. If the result drifted away from the person or the place you photographed, lower it. If the result looks like the photo with a filter on it, raise it. Strength and prompt push against each other, so a strong, specific prompt at low strength will usually be ignored.

Can you keep editing a result?

Yes. Refine takes the image currently on screen and feeds it back in as the new starting picture, so the second pass builds on the first instead of starting over from the photo. You can change the prompt, change the strength, or run the same settings again for a different roll of the dice.

Nothing is re-uploaded between passes because nothing was uploaded in the first place. The image is already in memory on the phone, so a refine pass costs the same as any other generation and works with no signal.

  • Walk a photo toward a look in small steps rather than one large jump, which tends to keep faces and geometry intact.
  • Fix a prompt that half worked: keep the result you liked, adjust the wording, and run a low-strength pass over it.
  • Every version is saved in the on-device gallery with its history, so an earlier pass is still there when the latest one goes wrong.
  • Refine works on every engine in the catalog, not only the free pack.

If you want to change one specific area and leave the rest of the picture untouched, that is masked inpainting, which is a separate tool and is available with the free fast model.

Which model pack should you use?

Start with the free one. Onhand is a free download and comes with a recommended fast model, roughly 2 GB, that handles image to image on its own. It is quick, it fits the smallest supported device, and for restyling work it is often all you need. The developer measured about 7 seconds for a 512 pixel image with this pack on an iPhone 14 Pro, with the method written up on the about page.

Pro is a one-time purchase that unlocks more model packs. There is no subscription, no credit system and no cap on how many images you make. The packs it adds give you more looks for portraits and scenes, and a higher-quality model that follows a longer, more specific prompt more closely. That matters most when your prompt is doing the heavy lifting at high strength.

PackCostWhat it adds for image to image
Fast Stable Diffusion class (SD Turbo), about 2 GBFreeQuick passes, masked inpainting, runs on the smallest supported device
XL class (SDXL Turbo)ProMore detail and a different set of looks for portraits and scenes
FLUX.2 Klein 4BProCloser prompt following on longer, more specific descriptions

Which packs your device can run depends on how much memory it has available, not on how fast the chip is. The free pack is the floor and runs on the developer’s test device, an iPhone 14 Pro with 6 GB. The larger packs need more room. Packs are between roughly 1 and 4.5 GB each and download once, after which the app works offline. Model licenses and credits are listed on the licenses page.

What can image to image not do here?

The honest list is short and worth reading before you plan a project around it. Onhand gives you a source photo, a prompt, a strength control and a refine loop. It does not give you the structural control rigs that desktop workflows use.

There is no ControlNet, no depth or edge conditioning and no pose guidance, so you cannot pin a skeleton or an outline and force the model to obey it. Strength is the only lever over how much structure is preserved. There is no LoRA support and no way to import your own checkpoint, so the range of looks is the range in the catalog. There is no outpainting, which means the result keeps the frame you started with, and no AI upscaling, so resize is an ordinary resize.

Text inside an image is the other thing to plan around. Diffusion models of this size get lettering wrong often enough that a sign, a label or a caption in the result should be treated as decoration, not as copy. Results are safe-for-work, and an on-device safety check runs over what comes out.

What it does not do

  • No ControlNet, depth maps or pose guidance. Strength is the only structure control.
  • No LoRA and no custom model import. The looks you get are the ones in the model catalog.
  • Text and lettering inside a generated image are unreliable and depend on the model. Do not count on a readable sign or caption.
  • No outpainting. The result stays inside the frame you started with.
  • No AI upscaling. Resize is a plain resize, not a detail-inventing one.

Frequently asked questions

Can I restyle a photo with AI on an iPhone without the internet?

Yes. Once a model pack has been downloaded, Onhand runs image to image entirely on the device GPU, including in airplane mode. The photo and the prompt never go to a generation server. You only need a connection to fetch a model pack in the first place.

What does the strength setting do in image to image?

It decides how much of your original photo survives the pass. Low strength keeps composition, faces and layout and mostly changes color and texture. High strength treats the photo as a rough guide and lets the prompt rebuild the scene. Start in the middle and move one step at a time.

Does image to image work with the free model?

Yes. Every model pack in the catalog supports image to image, including the free fast model that comes with the app. Pro packs are a one-time purchase and add more looks plus a model that follows longer prompts more closely. There is no subscription and no per-image credit.

Can I keep editing a result instead of starting over?

Yes, that is the Refine loop. It sends the image you are looking at back through as the new starting picture, so each pass builds on the last one. Versions are kept in the on-device gallery, so you can return to an earlier pass if a later one goes wrong.

Can I use ControlNet, a LoRA or my own model?

No. Onhand has no ControlNet, no depth or pose conditioning, no LoRA support and no custom model import. The only control over how much structure is preserved is the strength setting, and the available looks are the packs in the catalog. If a project depends on pose rigs or custom checkpoints, a desktop tool is the better fit.

Can I change just one part of a photo?

For a single area, use masked inpainting rather than image to image. You paint over the region you want redrawn and only that region is regenerated, which is available with the free fast model. Image to image reworks the whole frame every time, which is why strength matters so much.

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.