Skip to content
Beta open on TestFlight

Run frontier image models on your iPhone, not in a data center.

Private Diffusion is a local AI image generator for iPhone, iPad, and Mac. Every model is trained to hold its full-size quality and runs on your device's own chip, where nobody else sees the prompt.

Private Diffusion Studio on iPhone showing a finished image above the prompt and the model pickerPrivate Diffusion Studio on iPhone showing a finished image above the prompt and the model picker

Your Mac Becomes an Image Server for Claude Code

One click in the API Server tab and your Mac starts serving images to your own tools. Claude connects over MCP and generates with the models you have downloaded. Anything that speaks the OpenAI image API points at /v1/images/generations and gets the same, and every model but Nanosaur 2 streams partial images back as the picture forms if the client asks for them. Every image is made on your Mac and nothing passes through a cloud relay. The server answers on the network your Mac is on, and only to requests that carry your key.

Mac only. Text to image, on the models you have downloaded. Studio pauses while the server runs. On the Mac, images also drag straight into Finder and open in Quick Look, and the app sends a notification when a picture is done and puts Generate, Stop, and every tab on a keyboard shortcut.

Read the API docs
The API Server tab in Private Diffusion on a Mac, with the server runningThe API Server tab in Private Diffusion on a Mac, with the server running

Set Up With Your AI Agent

Paste this prompt into Claude Code or any coding agent. The agent checks that the API server on your Mac is running, then sets up MCP or a script in your own code.

What the Server Answers

  • GET /v1/models
  • POST /v1/images/generations
  • POST /mcp

Claude Code and Claude Desktop connect over MCP.

Bearer key, minted in the app.

Latent Upscaling to 6144 Pixels

Most upscalers stretch a finished picture. Private Diffusion enlarges the image before it is drawn, then draws it once, at full size, so fur, skin, and edges come out crisp instead of smeared. It runs on the Neural Engine in well under a second. No other image app on iPhone, iPad, or Mac does this. Set the size to extra large and the upscaler to 4x, and the long edge reaches 6144 pixels: 37.75 megapixels, a twenty-inch print at 300 dpi.

On Z-Image-Turbo, Juggernaut Z Fast, Chroma1-Flash, and Boogu Image Turbo, with a Mac that has 24 GB of memory. A 16 GB iPad reaches 4096 pixels on the same models.

From a Prompt to a Finished Picture on Your iPhone

You don't need an account. Write a prompt and watch your own device draw it.

  1. Private Diffusion Studio on iPhone with a prompt typed and the Generate button readyPrivate Diffusion Studio on iPhone with a prompt typed and the Generate button ready

    Describe It

    Type what you want to see, or tap the Surprise me! button and let the app suggest something. Pick a model and an aspect ratio, set a seed if you want the same picture twice, and tap Generate.

  2. Private Diffusion Studio on iPhone with the finished image the preview sharpened intoPrivate Diffusion Studio on iPhone with the finished image the preview sharpened into

    Watch It Form

    The picture sharpens on screen as your own chip draws it, and when it is done the prompt sits right under it, so you can change one word and go again.

    On the Mac, every model but Nanosaur 2 requests the Neural Engine for the preview while the diffusion runs on the GPU. Nanosaur 2 skips the preview and shows the finished picture.

  3. Private Diffusion image details on iPhone with the generation settings and the Recreate in Studio buttonPrivate Diffusion image details on iPhone with the generation settings and the Recreate in Studio button

    Keep the Recipe

    Every image you keep lands in your gallery with the model, prompt, seed, steps, and size that made it. Open one later, tap Recreate in Studio, and every setting comes back filled in.

How It Runs on Apple Silicon

Image models like Krea 2 Turbo usually run on servers you reach over the internet. Private Diffusion runs them on the chip already in your device.

  1. A Big Model, Made to Fit

    A full-size image model is too big for a phone or a tablet, so we shrink every model ourselves before it reaches your device. That is how Krea 2 Turbo, a 12 billion parameter model, runs on a 16 GB iPad or Mac.

  2. Trained So the Quality Holds

    Shrinking a model usually costs you detail. We train ours after we shrink them, against the full-size original, until the small one answers the same way.

  3. The Whole Chip, Put to Work

    Apple Silicon has more than one kind of processor, and Private Diffusion gives each one the work it does best. The GPU draws the picture for every model but Nanosaur 2, and on the Mac those models request the Neural Engine for the live preview, the final image, and the upscaler. On the newest Apple chips, the Neural Accelerators built into the GPU join in for all of them but Mage Flow Turbo and MOSS-Image. Nanosaur 2 runs entirely on Core ML, across the Neural Engine and the GPU.

  4. Private Because It Stays on Your Device

    Because your own chip does every step, your prompt and your picture never need a server.

A Private AI Image Generator That Runs on Your Device

Every step of the picture happens on the device in front of you. The app has no analytics SDK, and its privacy manifest declares that it neither tracks you nor collects data.

Private Diffusion Gallery on iPhone, with Face ID unlocking the Hidden tabPrivate Diffusion Gallery on iPhone, with Face ID unlocking the Hidden tab
Incognito Mode
Turn it on in Studio and nothing you make is written to the gallery. On iPhone and iPad it also blanks the app-switcher preview, so a glance at your open apps shows neither the image nor the prompt.
A Hidden Tab With a Lock
Hide an image and it moves to the gallery's Hidden tab, which asks for Face ID, Touch ID, or your passcode (your login password on a Mac) before it opens. That lock is on from the first launch.
Your Own iCloud
If you use iCloud, your gallery syncs through your own private database, which we cannot see. Signed out, it stays on the device.
Works in Airplane Mode
Once a model is downloaded, the app needs no connection at all. Switch on Airplane Mode and it keeps making pictures.

What Other Image Apps Say They Collect

Private Diffusion never sends us a prompt or an image. The table lists what other image apps, on the device and in the cloud, say they collect.

AppWhat the Developer Discloses
Draw ThingsData Not Linked to YouPurchases, User Content, Identifiers
ComfyUIComfyUI's desktop app, Comfy Desktop, sends usage and error telemetry to PostHog and Datadog, and its first-launch telemetry box comes checked.
KreaData Linked to YouContact Info, User Content, Identifiers, Usage Data
Adobe FireflyData Linked to YouPurchases, Contact Info, User Content, Identifiers, Usage Data, Diagnostics
ChatGPTData Linked to YouHealth & Fitness, Location, Contact Info, User Content, Search History, Identifiers, Usage Data, Diagnostics
Leonardo.AiData Linked to YouContact Info, User Content, Identifiers, Usage Data

App Store labels as each developer declared them on the US App Store, and the privacy policy and code of Comfy Desktop 1.1.6, all read October 11, 2026.

One Download, Then Start Making Pictures

Get the app, pick a model, and make a picture. We convert and train every model for Apple chips before it reaches you.

The Model Library tab in Private Diffusion on iPhone, with one model mid-downloadThe Model Library tab in Private Diffusion on iPhone, with one model mid-download
  • Getting started

    Private Diffusion
    Download a model, type a prompt
    Draw Things
    Pick a model, then tune the settings
    ComfyUI
    Python install, then a node graph
  • How models are shrunk

    Private Diffusion
    Retrained after shrinking (QAD)
    Draw Things
    Compressed after training, no retraining
    ComfyUI
    Depends on the checkpoint you load
  • Processors it uses

    Private Diffusion
    GPU and Neural Engine
    Draw Things
    GPU; Neural Engine for select models
    ComfyUI
    GPU through PyTorch MPS
  • Data the app collects

    Private Diffusion
    None
    Draw Things
    Purchases, User Content, Identifiers
    ComfyUI
    Usage and error telemetry, on by default

Full comparison with Draw ThingsFull comparison with ComfyUI

An AI Image Prompt Generator That Shows What Each Word Does

Choose a light, a camera, a style. Every choice is a picture made on the same seed, so you see what one word changes before you use it. Then copy a prompt that works in Private Diffusion or any other image generator. What you type stays in your browser.

  • a fox in a misty forest
    a fox in a misty forest
  • a fox in a misty forest, with golden hour added
    + Golden hour
  • a fox in a misty forest, with moonlight added
    + Moonlight
  • a fox in a misty forest, with watercolor added
    + Watercolor
Open the prompt generator

Which iPhone, iPad, or Mac Runs a Local AI Image Generator?

We convert, quantize, and publish every model ourselves, one download each. A 16 GB iPad or Mac runs every one of them, and the table shows which ones your iPhone runs.

ModelDownloadiPhoneiPadMac
AnimaAnime and illustration styles on the lightest tier, with a steps slider to trade speed for detail.2.51 GB8 GB8 GBAny
Flux.2 Klein 4BFour steps on every supported device, with negative prompts and Rewrite Prompt.4.94 GB8 GB8 GBAny
Nanosaur 2The one to start with: the smallest download, on every supported device.1.14 GB8 GB8 GBAny
Mage Flow TurboFlagship-quality images in four steps on a 12 GB iPhone or iPad, or a 16 GB Mac.7.88 GB12 GB12 GB16 GB
MOSS-ImageShort text inside the picture, on a sign, a poster, or a label, in English, Chinese, Japanese, and more.11.01 GB12 GB12 GB16 GB
Krea 2 TurboHigh-fidelity images from a 12B model, in eight steps, or in four or two with an adapter on devices with 18 GB of memory or more.12.09 GBNot available on iPhone16 GB16 GB
Kroma TurboKrea 2 Turbo with classifier-free guidance, for tighter prompt adherence.12.08 GBNot available on iPhone16 GB16 GB
Z-Image-TurboNoticeably finer detail, with 4x upscaling on devices with 12 GB of memory or more.6.40 GB12 GB8 GBAny
Juggernaut Z FastRunDiffusion's photorealistic Z-Image fine-tune. Pick it when you want an image that reads as a photograph rather than a render.6.29 GB12 GB8 GBAny
ERNIE Image TurboFollows long, specific prompts, with an on-device prompt enhancer.9.66 GBNot available on iPhone16 GB16 GB
Chroma1-FlashNegative prompts, and a latent upscaler that reaches 4x.10.34 GBNot available on iPhone16 GB16 GB
Boogu Image TurboA large model that finishes in four steps.12.86 GBNot available on iPhone16 GB16 GB

Requires iOS, iPadOS, or macOS 26.4 on an iPhone 15 Pro or later, an iPad with an M-series chip, or any Apple Silicon Mac.

Read about each model

Be First in Line

Pre-order now, and Private Diffusion downloads itself to your device the day it launches.

Pre-order on the App Store
Private Diffusion Studio on a Mac with a finished Krea 2 Turbo imagePrivate Diffusion Studio on a Mac with a finished Krea 2 Turbo image
Private Diffusion Studio on iPad in landscape, running Krea 2 TurboPrivate Diffusion Studio on iPad in landscape, running Krea 2 Turbo
Private Diffusion Studio on iPhone with a finished imagePrivate Diffusion Studio on iPhone with a finished image