What is alis-studio-mobile?

Alis Studio Mobile is an iPhone app that turns text prompts into images entirely on the phone itself, with no internet connection needed after the first setup and no cloud service or account involved. It uses a technology called MLX Swift, made by Apple's machine learning team, to run an image generating AI model directly on the device's own chip. It is the phone version of an existing Mac app called Alis Studio, and shares that app's look and color scheme. Because generating images this way requires loading a multi gigabyte AI model fully into memory, the app only works on iPhones with at least 8 gigabytes of RAM, roughly the iPhone 15 Pro and newer higher end models. Older or smaller phones are blocked from generating images and shown an unsupported device message instead, since trying anyway would cause the phone's operating system to shut the app down partway through. The app also only runs on a real iPhone, not in Apple's simulator on a Mac, because the underlying technology needs a real device's graphics hardware. The first time you open the app it needs an internet connection to download the AI model itself from Hugging Face, a repository for AI models, which can be several gigabytes and takes a while, after which the model is stored on the phone and later launches are fast. You can choose between two versions of the model that trade some image quality for speed, adjust how many generation steps to run, cancel a generation in progress, and browse a gallery of past images saved on the phone. There is no App Store version. Installing it means building the app yourself using Xcode, Apple's developer tool, on a Mac, then loading it onto your iPhone with a free Apple account, no paid developer subscription required. Apps installed this way with a free account expire after seven days and need to be reinstalled from the Mac to keep working, and a free account can only have three such apps installed at once. The project is released under the MIT license, and part of its underlying image generation code is adapted from an example project published by Apple, also under the MIT license.