# Answers about the local AI image generator on your iPhone, iPad, and Mac. Which devices run which models, what you can set besides the prompt, where your pictures end up, and what the app sends us, which is nothing. Where the answer is no, or not yet, it says so. [Getting Started](#faq-gettingStarted)[Creating](#faq-creating)[Privacy](#faq-privacy)[Sharing Your Images](#faq-sharing) ## Getting Started With the Local AI Image Generator - ### Which devices can run Private Diffusion? You need iOS, iPadOS, or macOS 26.4 or later and an Apple Silicon chip: an iPhone 15 Pro or later; the iPhone 17 Pro, 17 Pro Max, and iPhone Air have 12 GB and run three more models. On iPad you need an M-series chip, and only an iPad Pro with 1 TB or 2 TB of storage has the 16 GB that runs all ten. Any Apple Silicon Mac works. Intel Macs are out, and so is any device under 8 GB, which is the floor the smallest model is gated to. - ### Does it work without an internet connection? Yes. A model downloads once, and after that Private Diffusion is an offline AI image generator. Prompts, previews, and finished images all happen on the device, with no sign-in and no server on the other end. - ### How much storage do the models need? Between about 2.5 GB (Anima) and 13 GB (Boogu Image Turbo) per model. Models that share a text encoder cost less together: with Flux.2 Klein 4B installed, Z-Image-Turbo adds about 4 GB instead of 6.4, and with Krea 2 Turbo installed, Kroma Turbo adds about 7.9 GB instead of 12. Downloads keep going with the app closed and resume on their own if they stall, and you can delete any model to reclaim space and fetch it again later. - ### Why can't my iPhone run every model? Memory decides. Anima and Flux.2 Klein 4B run on 8 GB devices. Mage Flow Turbo, Z-Image-Turbo, and Juggernaut Z Fast need a 12 GB iPhone, though the two Z-Image models also run on an 8 GB iPad. Krea 2 Turbo, Kroma Turbo, ERNIE Image Turbo, Chroma1-Flash, and Boogu Image Turbo need an iPad or Mac with 16 GB or more, and are not offered on iPhone at all. - ### What does the Mac version add? Menu bar commands and keyboard shortcuts, a window per image, and a local image API and MCP server you can switch on for your own tools. The API is off until you turn it on, and it answers on the network your Mac is on, only to requests that carry your key. - ### How is this different from Draw Things? Draw Things does more: checkpoints, samplers, config files, and hundreds of models. Private Diffusion is what you open when you want the picture, not the pipeline. Three differences: - It uses both processors. The diffusion model runs on the GPU through Metal, and the Neural Engine takes the live preview, the final decode, and the upscaler. A few of the largest decodes come out cleaner on the CPU, so those go to the CPU. You never choose, and there is no setting to go find. - Every model is quantization aware distilled. We train the weights after we shrink them, against the full size original, until the small one answers the same way. The picture you get is the one the full size model would have made. - Private Diffusion's privacy manifest declares zero collected data types. Draw Things' App Store listing discloses Purchases, User Content, and Identifiers as collected but not linked to your identity. ![App Store App Privacy card for Draw Things, listing Purchases, User Content, and Identifiers as data collected but not linked to your identity.](/_astro/draw-things-app-privacy.BwwWhgXv_ZeFPjW.png) Draw Things' own App Store privacy disclosure, as of 30 August 2026. - ### Can I buy it before launch? Yes. Pre-order on the App Store and it downloads itself to your device on release day. - ### What will it cost? We have not announced pricing yet. We would rather tell you that than guess in public. ## Creating - ### What can I control besides the prompt? Seven aspect ratios, three output sizes on most devices plus an extra-large tier on high-memory ones, and a seed on every model. An upscaler reaches 4x on four models and 2x on the rest. Negative prompts work on five models, the on-device Rewrite Prompt enhancer on four, and Anima and Kroma Turbo add a steps slider. Each model page lists exactly what it supports. - ### Does it do image-to-image or inpainting? No. Private Diffusion turns text into images, and that is the whole product for now. There is no image-to-image, inpainting, or ControlNet, and we would rather say so here than have you find out after the download. ## Privacy - ### Is anything sent to your servers? No. Apart from iCloud sync for your gallery and a few settings, the only outbound calls Private Diffusion makes are to Apple's CDN, to download a model. After that there is nothing to send and nowhere to send it: no ads, no analytics, no tracking, no crash reporter. The privacy manifest declares zero collected data types. - ### Where are my images stored? On your device. If you are signed in to iCloud, your gallery also mirrors into your own private iCloud database, the way Notes does. Numen has no server in that path and no way to read it. Sign out of iCloud and it stays on the one device. - ### How do I keep images private on my device? Three layers. Incognito Mode keeps new generations out of your gallery entirely, and on iPhone and iPad it also covers the screen in the app switcher. The Hidden tab asks for Face ID, Touch ID, or your device passcode before it opens, and that lock is on from the first launch. And you can hide the Hidden tab itself. - ### Does the app filter or censor what I create? Generation happens on your device, so no cloud service reviews your prompts and there is no account to suspend. If your device's system-level Sensitive Content Warning is on, images it flags are blurred until you tap Show. Nothing is blocked or deleted, and that switch lives in iOS and macOS Settings, not in the app. ## Sharing Your Images - ### Why do exported images have a small watermark? Exports carry a small Private Diffusion wordmark in the bottom-right corner. It goes on every way an image leaves the app: saved to Photos, shared, copied, dragged out, or exported to a file on a Mac. What you see in the gallery is always clean. It is on by default, the switch is in Settings under Exports, and your choice follows you to your other devices. Images the system has marked sensitive export without it either way. - ### Do exported images include my prompt? Yes. Exports embed the prompt, model, seed, and settings as IPTC provenance metadata and mark the file as AI-generated, so an image can always be traced back to how it was made and recreated. Worth knowing before you post one publicly. - ### Can I use my images commercially? For most of them, nothing stands in your way. Six models ship under Apache 2.0, which licenses the model and never mentions the pictures, so there is no license term for your images to trip over. Three carry conditions. Krea 2 Turbo and Kroma Turbo, which is built on it, fall under the Krea 2 Community License: commercial use while your company-wide revenue over the trailing twelve months stays under one million US dollars, and an enterprise license from Krea at or above that. Anima asks for an NVIDIA attribution, which is on its model page, and does not let you use its output to train a model that competes with it. Mage Flow Turbo has no license on record, so we make no promise for it. Each model page lists what applies. [![Pre-order on the App Store](/app-store/pre-order-badge/en/pre-order.svg)![Pre-order on the App Store](/app-store/pre-order-badge/en/pre-order-dark.svg)](https://apps.apple.com/app/apple-store/id6760593556?pt=96139298&ct=pd_faq_page&mt=8)