All features

Offline generation

An offline AI image generator for iPhone

Onhand by ArkalEdge runs image generation and editing on the iPhone's own GPU, with no cloud upload. You download a model pack once, and after that you can create on a plane, on the underground, or anywhere the signal drops.

By Onhand · Updated September 30, 2026

What you get

  • Text-to-image generation with no network connection, once a model pack is on the device
  • Image-to-image restyling of a photo from your library, with a strength control, on every model
  • A Refine loop that keeps iterating on a result and keeps the earlier versions
  • Masked inpainting: paint over an area and regenerate just that part, with the free fast model
  • Resize and on-device background removal for photos you bring in
  • Style presets and an on-device gallery with history, all readable offline
  • An on-device safe-for-work check that runs on results without a connection

How does offline generation work?

Onhand by ArkalEdge generates images on the iPhone itself, with no cloud upload, so it works in airplane mode after a one-time model download. There is no rendering server in the loop and no queue to wait in. The prompt you type, the photo you pick and the image that comes out stay on the phone.

The one-time part is the model pack. An image model is a large file of learned weights, and the weights have to be on the device before the device can use them. The first time you choose a model in Onhand, the app downloads that pack and stores it. That download is the only step in making a picture that needs a network, and you do it once per pack rather than once per image.

After the pack is in place, the work happens on the graphics processor built into the phone. The app turns your words into numbers the model understands, the GPU runs the denoising steps that turn random noise into a picture, and the finished image is written to a gallery inside the app. Nothing in that sequence reaches out to a server, which is why pulling the network away changes nothing about it.

  1. Download a model pack once, over Wi-Fi, and wait for it to say it is ready.
  2. Type a prompt, or bring in a photo from your library to restyle.
  3. The phone's GPU runs the model and builds the image locally.
  4. The result lands in the app's own gallery on the device.
  5. Refine it, restyle it again, or paint over part of it and regenerate, still with no connection.

Editing works the same way as generating, because it is the same machinery. Image-to-image feeds your photo into the model instead of starting from noise, and a strength control decides how far the result travels from the original. Inpainting limits that work to the area you paint. Both run on the device, so both run offline.

What do you need to run it offline?

You need an iPhone or iPad on iOS 26 or iPadOS 26 or later, free storage for at least one model pack, and enough memory to hold the pack you want to use. Onhand is a free download and a recommended fast model is free to use, so the smallest useful setup is one free pack and the space to keep it.

Model packs run roughly 1 to 4.5 GB each. The free pack is a fast Stable Diffusion-class model of about 2 GB, which is the one most people start with. Pro packs are larger: they add more looks for portraits and scenes, and a higher-quality model with better prompt following. Pro is a one-time purchase that unlocks those packs, with no subscription, no credits and no limit on how many images you make.

Memory is the constraint that decides which packs a device can run, not how new the chip is. A model has to fit in memory in one piece while it works, so a phone with more memory can run larger packs and a phone with less cannot, whatever its benchmark scores say. The developer's test device and floor is an iPhone 14 Pro with 6 GB of memory, and the free pack runs there. Larger packs need more headroom than that.

  • iOS 26 or iPadOS 26 or later, on an iPhone or iPad.
  • Free storage of roughly 1 to 4.5 GB for each model pack you keep.
  • A connection once, for the pack download. Wi-Fi is the sensible choice at that size.
  • Enough device memory for the pack. Memory decides the ceiling, not chip speed.
  • No account and no sign-in at any point.

What still needs a connection?

Generation, editing and your gallery work with the network off. Getting the app and getting a model pack do not. That is the whole split, and it is worth stating precisely rather than just saying the word offline.

TaskWorks with no connection
Installing the appNo, the App Store needs a connection
Downloading a model packNo, this is the one-time download
Text-to-image generationYes, once the pack is on the device
Image-to-image and the Refine loopYes
Masked inpainting with the free fast modelYes
Resize and background removalYes
Browsing and exporting from the galleryYes
On-device safety check on resultsYes
Buying Pro or restoring a purchaseNo, Apple's purchase service needs a connection
Seeing newly added model packs in the catalogNo, the catalog list is fetched

One more thing belongs in the honest column. Onhand uses limited anonymous analytics and Apple and RevenueCat purchase services, and those do talk to the internet when the phone is online. Your prompts, your source photos and your generated images are not part of that traffic. The full account of what is collected is on the privacy page.

What should you expect in speed and quality?

Expect a picture in seconds, at a smaller size and a rougher finish than a large cloud service gives you. The developer measured about 7 seconds for a 512 pixel image with the free fast model and about 12 seconds with the XL pack on an iPhone 14 Pro; the method is described on the about page. Newer devices with more memory have more room to work with.

The quality gap is real and it comes from physics rather than from effort. A phone has a small fraction of the memory and power of a rented datacentre GPU, so the models that fit on it are smaller models. Fine detail, crowded scenes with several subjects and precise composition are where that shows first. The larger Pro model follows a prompt more closely than the fast free one, which is most of what you are paying for.

Text inside images is the weak point across this whole class of model. Signs, labels and logos tend to come out as letter-shaped marks rather than words you can read. Plan on adding real type yourself afterwards and treat any legible lettering as luck.

Negative prompts also do nothing on the Turbo models, which run without guidance. Put the style you want into the positive prompt instead of listing what you want to avoid, and use the Refine loop to walk a result closer to what you had in mind.

How can you check it is really offline?

Turn on airplane mode and generate an image. That is the entire test, it takes a minute, and it is worth doing once so the claim is yours rather than a line in marketing copy. An app that quietly needs a server fails this immediately.

  1. On Wi-Fi, download a model pack and wait until the model shows as ready.
  2. Open Control Centre, switch on airplane mode, and check that Wi-Fi is off as well.
  3. Type a prompt and generate. The image should appear in the usual few seconds.
  4. Pick a photo from your library and restyle it, then run Refine on the result.
  5. Open the gallery and confirm your earlier images are all still there.

It is also the clearest answer to the privacy question, because a prompt that never leaves the phone cannot be logged, trained on or leaked by a service you do not control. Offline capability and privacy are not identical claims, and working in airplane mode is strong evidence that the work happens where the app says it does.

What it does not do

  • No cloud fallback. If a model pack is not on the device, there is nothing to fall back to
  • No custom model import, no LoRA and no ControlNet. You use the packs in the catalog
  • No AI upscaling and no outpainting. Resize is a plain resize, and the canvas does not extend
  • No video, and no Mac app. Onhand is an iPhone and iPad app
  • Text and lettering inside images are unreliable on these models and are never promised

Frequently asked questions

Can an AI image generator work without an internet connection?

Yes, if the model runs on your device instead of on a server, with no cloud upload for generation. The model file has to be downloaded once, and after that the phone does the computing itself. Onhand by ArkalEdge works this way, so generation and editing continue in airplane mode. Services that render in the cloud cannot do this, because the machine doing the work is somewhere else.

Does Onhand need a connection the first time?

Yes, twice. You need a connection to install the app from the App Store, and a connection to download the model pack you want to use, which is between roughly 1 and 4.5 GB. Once a pack is on the device, making images with it needs nothing further.

How much storage does an offline AI image generator need?

Plan on roughly 1 to 4.5 GB for each model pack you keep. The free fast model is about 2 GB. You can keep several packs if you have the space, or remove one you are not using and download it again later.

Will it run on my iPhone?

It needs iOS 26 or later, and the pack you can run depends on how much memory the device has rather than how fast the chip is. The free pack runs on the developer's test device, an iPhone 14 Pro with 6 GB of memory, which is the floor the app is built against. Larger packs need more memory than that.

Is generating images offline the same as generating them privately?

They overlap without being the same claim. Running offline means your prompts and photos are not sent to a generation server, which is the core of the privacy argument. Onhand still uses limited anonymous analytics and Apple's purchase services when the phone is online, and the privacy page sets out exactly what that covers.

Can I edit my own photos offline, or only generate new images?

Both. Image-to-image restyling works on every model with a strength control, masked inpainting works with the free fast model, and resize and background removal run on the device as well. All of it works with the network off, and the results go into the app gallery on the phone.

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.