Offline AI Image Generator for iPhone and Android

Yes, a recent phone can generate AI images offline. On iPhone and iPad, Draw Things runs Stable Diffusion-style models locally for free, and Apple’s Image Playground creates images in a few set styles on Apple Intelligence devices. On Android, the open-source Local Dream app runs Stable Diffusion on the phone and uses the NPU on supported Snapdragon chips. Expect slower results and smaller models than cloud generators like ChatGPT or Midjourney, and plan for several gigabytes of storage. What you get in return is that your prompts and images never leave the phone.

Which apps generate AI images offline? #

Draw Things (iPhone and iPad) #

Draw Things describes itself as a way to “generate images locally and offline,” and says its free edition “runs offline on all devices to protect your privacy.” You download image models inside the app, then generate without a connection. It can also train LoRAs, small add-ons that teach a model a style or subject, entirely on the device. It’s on the App Store for iOS, iPadOS and macOS, with a paid tier for extra features. See the Draw Things site for current details.

Image Playground (Apple Intelligence) #

Apple’s Image Playground makes images in styles such as Animation and Illustration, from a description or a photo. It’s part of Apple Intelligence, so it needs a supported device: iPhone 15 Pro and later, and iPads and Macs with M1 or newer chips. Apple describes it as running on the device. It’s simple and fun, but deliberately limited: no photorealism, and no choice of model. Our guide to whether Apple Intelligence works offline covers which of its features need the cloud.

Local Dream (Android) #

Local Dream is an open-source Android app for “Android Stable Diffusion with Snapdragon NPU acceleration,” which also supports CPU and GPU. Per its documentation, NPU support for Stable Diffusion 1.5 needs a Snapdragon chip with a Hexagon V68 NPU or newer, and SDXL needs a Snapdragon 8 Gen 3 or newer. On other chips, it can only run Stable Diffusion 1.5 on the CPU or GPU, which is slower.

Why does the NPU matter for images but not for chat? #

This is the one place where a phone’s AI accelerator, the NPU, clearly earns its keep.

Image generation with a diffusion model is compute-heavy. The model starts from noise and refines the whole image over many steps, doing a huge number of calculations each time. That’s exactly the kind of work NPUs are built for, and it’s why chip makers advertise their performance in TOPS, trillions of operations per second.

Text generation in a chatbot is different. It’s mostly limited by how fast the phone can read the model from memory, one word at a time, so raw compute matters less and RAM matters more. That’s why most local chat apps run on the GPU or CPU and skip the NPU entirely. We explain that side in do you need an NPU to run AI.

So if offline image generation is your main goal, a phone with a strong, well-supported NPU or GPU matters. For offline chat, prioritize RAM, as covered in how much RAM you need to run an LLM on a phone.

What to expect from on-device image generation #

• Speed. Each image can take from several seconds to minutes, depending on the model, the image size, the number of steps and your chip.
• Storage. Image models are large. Budget several gigabytes for one or two models.
• Heat and battery. Generating a batch of images works the chip hard. Phones warm up and slow themselves down in long sessions.
• Quality. Older, smaller models like Stable Diffusion 1.5 are less polished than today’s cloud generators, especially with hands, text in images and complex scenes.
• Setup. Apps like Draw Things have many settings. The defaults are a fine place to start.

Cloud vs offline image generation #

Cloud tools keep your prompts and the images they make under their privacy policies. That’s fine for most uses, and less fine for a photo of your child you want restyled, or images for a project that isn’t public yet. For more on what happens to uploads, see is it safe to upload photos to ChatGPT.

Whatever tool you use, don’t generate images of real people without their consent, and check the rules before using generated images commercially.

Where a local chat app fits #

Personal LLM is not an image generator. It runs language models on your phone, and every model in its catalog can take an image as input once you add the image-support file. That makes it the other half of the picture: it can describe a photo, read a sign or a form, or explain a screenshot, all offline. See ask AI about a photo offline for what that’s good for.

It’s also handy for writing better prompts for an image generator without a connection:

Write a detailed image prompt for a cozy reading nook in a treehouse at dusk. Include lighting, camera angle, style and color palette, under 60 words.

Frequently asked questions #

Can I generate AI images offline on my phone? #

Yes. Draw Things on iPhone and iPad and Local Dream on Android run image models on the device once you’ve downloaded them. Apple Intelligence devices also have Image Playground built in, for images in set styles.

Is Draw Things free? #

It has a free edition that generates images locally and offline, plus a paid tier with extra features. Check the App Store listing for current pricing.

Can Android phones run Stable Diffusion? #

Yes. Local Dream runs Stable Diffusion on Android, using the NPU on supported Snapdragon chips and the CPU or GPU on others. Newer Snapdragon flagships can run larger models like SDXL.

Do I need an NPU to generate AI images on a phone? #

Not strictly, but it helps a lot. Image generation is compute-heavy, so an NPU or a strong GPU makes each image much faster. Without one, generation falls back to slower paths and smaller models.

Can ChatGPT or Gemini generate images offline? #

No. Their image generation runs on the companies’ servers and needs a connection. For offline images, use an on-device app.