All articles

Guides

Will AI Image Generation Run on My iPhone?

Which iPhones can run on-device diffusion models, why available memory rather than chip speed decides it, and what to expect on older hardware.

By · · Updated · 8 min read

Whether AI image generation runs on your iPhone depends mostly on available memory, not chip speed. For Onhand, first check iOS or iPadOS 26 support and the model list shown on your device; the app uses a 6 GB minimum memory tier for its current catalog. More memory can provide headroom, but neither 6 GB nor 8 GB alone establishes that every model and editing task will succeed.

The deciding factor is available memory, rather than chip speed or free storage. Every other app works the opposite way: if it downloads, it runs, and a faster phone runs it better. An AI image generator breaks both of those rules. (Programs of this kind are called diffusion models.) It has to hold billions of numbers in memory all at once before it can do anything, so memory is a hard gate. Speed only decides how long you wait once you are past it.

The rest of this post covers how that gate works and what each memory tier means in practice. It also covers how to check where your phone sits, before you spend a few gigabytes finding out.

Why memory is the deciding factor, not speed

The active model components and their temporary data must fit the memory available to the app. Pipelines can load components in stages, but each stage still has a working set it must accommodate. File size alone does not reveal that peak, and memory pressure can change during a run.

Chip speed changes a different number. A faster graphics chip (a GPU) does the same math in less time. So an image that takes twelve seconds on one phone might take twenty on an older one. That is a worse experience, not a broken one. A model that needs more working room than the system will grant does not run slowly. It does not run.

Free storage is not memory, and iOS enforces the difference

A phone can have 200 GB free and still fail to run a 3 GB model. Storage and memory are two separate resources. Storage holds the model file while it sits on disk doing nothing. Memory is the much smaller working space the model must be unpacked into before it can compute anything. (Memory is also called RAM.) Downloading is a storage question and running is a memory question; passing the first tells you nothing about the second.

On a desktop computer the difference is softer. When a desktop runs short on memory, it parks some of that memory on the disk instead. (The technical names for that are paging and swap.) Everything slows to a crawl, but the job finishes. iOS has no swap file for app memory. It also caps how much memory any single app may hold. That cap sits well below the total RAM of the device, because the system and everything else need their share. An app that crosses the ceiling is shut down on the spot, without warning. From the outside, that looks like the app vanishing in the middle of making your image.

This is why a well-built on-device app checks the memory budget before it offers you a model. The alternative is letting you download three gigabytes and find out the hard way. We go through how that budget works in why RAM decides which models your iPhone can run.

Gate 1: free storagemodel file, plenty of roomGate 2: memory iOS grants the appweights + working setDevice RAMPer-app limitrest reserved for the system
Two gates. Almost any phone lets a model through the storage gate. The second gate is much narrower: the memory iOS will grant a single app.

What each memory tier can realistically run

Higher-memory devices can have more room for demanding models, but a 2 GB or 3 GB download is not a universal compatibility class. Compression, engine, output size, and loading strategy all change the memory needed. Use the model’s actual requirements and the app’s eligibility checks.

  • 8 GB and above. More RAM can provide useful headroom. Check the operating system and per-model requirements; it is not a blanket approval for every model.
  • 6 GB. The smaller compressed models are the reliable choice. Larger ones depend on how the app manages memory. That means loading the chain of steps (the pipeline) in stages rather than all at once. It also means building the final image in small tiles instead of one full-resolution buffer.
  • 3–4 GB. Below Onhand’s current minimum model tier. That is a statement about this app’s catalog, not proof that no image model of any kind can run on such a device.

Two things shift the line. The first is compression: storing each learned number in 4 or 8 bits instead of 16. That is what brings a multi-gigabyte model down to phone size at all. (The technical name is quantization.) The second is that a model needs more room than its own numbers. The scratch data in between goes by the name intermediate tensors, and it needs space. So does the final decode, the step that turns that data into a picture, which is where memory use peaks. The question is whether the whole chain of steps fits under the cap at its peak moment.

What degrades on older hardware, and what will not run

On an older phone with enough memory, generation works and is slower. Each denoising step takes more time, and the final image takes longer to assemble. Heat becomes a factor sooner, too. The chip lowers its own speed to shed heat, so the fifth image in a row takes longer than the first. (The technical name for that is thermal throttling.) None of this is failure; the output is the same, later.

If a task exceeds the available budget, it may fail or the system may terminate the app. A smaller output, a different model, or a pipeline with lower memory use can help where supported. The current catalog allows Detailed Images on the 6 GB tier, but that does not establish a fixed completion time. Try the actual workflow you need.

How to tell before you download anything

Check two things. First, how much memory your phone has. Second, what the app says about the specific model. Your iPhone does not list its RAM anywhere in Settings. It is a published spec for every iPhone model, though. Look up your exact model, then place that number against the tiers above.

  1. Find your RAM. Settings shows free storage, and that is the wrong number here.
  2. Look for a per-model compatibility signal. A serious on-device app marks which models your device can run. A 3 GB download offered with no mention of memory is a warning sign.
  3. Start with the smaller model. It answers the question for two gigabytes of download instead of three.

Do not judge a phone by how new it feels. Memory does not track release year cleanly across the product line. A recent phone with less memory can be a worse fit than an older one with more.

Where Onhand sits on this

Onhand requires iOS or iPadOS 26 and checks the device’s memory tier against the model catalog. The current catalog gives its listed models a 6 GB minimum tier, including Detailed Images and Detailed Scenes. This is an eligibility rule, not proof that every model and editing route has been benchmarked on every 6 GB device.

Use the Models screen as the practical check, install the recommended Turbo pack, and run the workflow you need before downloading more. Model downloads range from roughly 1 GB to 4.6 GB. Allow space for installation and your gallery as well. The Stable Diffusion on iPhone guide explains why download size and working memory differ.

On an iPad, the same memory check applies; see the AI image generator for iPad. If your iPhone cannot run iOS 26, Onhand vs LocalGen covers an app that runs on iOS 18 and recommends iPhone 12 or later.

Sources

Primary sources for the claims above, so you can check them rather than take our word for it.