Skip to content

well9472 · 670M · Nanosaur2-670M

Run Nanosaur 2 on iPhone, iPad, and Mac: Illustration in a 1.14 GB Download

Nanosaur 2 is an illustration model from well9472 that reads tags or plain sentences. Private Diffusion runs it entirely on Core ML, on every iPhone, iPad, and Mac the app supports. Its images stop at the L size, because it offers neither an XL size nor upscaling.

Parameters
670M
License
MIT License
Download
1.14 GB

View on Hugging Face

Nanosaur 2 runs in Private Diffusion, a local AI image generator for iPhone, iPad, and Mac. Every model in the app is trained with quantization aware distillation (QAD) to stay near-lossless at a fraction of its full size. Download it once inside the app.

Pre-order on the App Store
Nanosaur 2 in Private Diffusion Studio on iPhone, with a finished image

Nanosaur 2 Requirements: Will It Run on My Device?

  • iPhone

    Runs on every supported iPhone

    iPhone 15 Pro and later, from 8 GB of memory up.

  • iPad

    Runs on every supported iPad

    Any iPad with an M-series chip.

  • Mac

    Runs on any Apple Silicon Mac

    M1 or later, from 8 GB of memory up.

Check your device

Nanosaur 2 Specs

Download

Download Size
1.14 GB

Model

Maker
well9472
Based On
Nanosaur2-670M
Parameters
670M
Quantization
Trained 8-bit (QAD)
License
MIT LicenseCovers the image model and its VAE, not the pictures you make with them. Nothing in it asks anything of your images.

Generation

Steps
4 (fixed)
Upscaler
None
Negative Prompt
No
Rewrite Prompt
No
Output Sizes
  1. S · 512 pxEvery device
  2. M · 768 pxEvery device
  3. L · 1024 pxEvery device

Nanosaur 2 Examples and Prompts

25 of 25 prompts

Browse Illustration prompts

What Nanosaur 2 Is Good At

It learned from about four million illustrations, and illustration is what it makes. We ship it at 8-bit, because a model this small loses too much quality at a lower bit width.

Nanosaur 2's text encoder is Google's Gemma 3 270M, modified by Numen. Gemma Terms of Use

Pick It For

  • Illustrated characters and scenes, written as tags or as plain sentences
  • A 1.14 GB download, the smallest in the app, that runs on an 8 GB iPhone
  • Control over emphasis: wrap a word in a weight, such as (fennec ears:1.3), and it counts for more

Know Before You Download

  • Images stop at the L size, with neither an XL size nor an upscaler
  • Fixed at four steps, and it offers neither a negative prompt nor Rewrite Prompt
  • The first image at each size can take a few minutes while that size is prepared, and sizes are prepared again after each app update
  • On disk, it can use up to 3 GB on iPhone, or 4 GB on iPad or Mac, as sizes are prepared
  • On iPad, leaving the app stops the image it is making

Nanosaur 2 in Private Diffusion vs. Draw Things vs. ComfyUI

Nanosaur 2 is the same model wherever you run it. The apps differ in how they shrink it to fit, which chips do the work, and who sees your prompts.

Private Diffusion compared with Draw Things and ComfyUI
FeaturePrivate DiffusionDraw ThingsComfyUI
Image QualityNear-lossless: every model is quantization aware distilled (QAD) against the full-size originalShrinks models after training (post-training palettization)Runs the weights you load
SpeedCore ML across the Neural Engine and the GPUGPU; the Neural Engine only for i8x models on M3 and M4PyTorch MPS, the slowest inference framework on Apple platforms
PrivacyCollects no dataPrivacy label lists Purchases, Identifiers, and User ContentSends telemetry to PostHog and Datadog, checked by default on first launch
SecurityApp Sandbox and Hardened RuntimeApp Sandbox (Mac App Store)Unsandboxed Electron
The AppNative Swift with Liquid Glass, macOS 26.4 or laterMac CatalystElectron

How to Run Nanosaur 2 on iPhone, iPad, and Mac

  1. Pre-order Private Diffusion on the App Store; it downloads itself on release day.
  2. Open Downloads in the app and tap Nanosaur 2. It is a single 1.14 GB download with nothing to configure.
  3. Pick Nanosaur 2 in Studio, describe the picture, and generate. It runs on your own device, so there is no queue.
Pre-order on the App Store

Questions About Nanosaur 2

  • Yes. Every iPhone from the 15 Pro onward has at least 8 GB of memory, which is what Nanosaur 2 needs, so it runs on every supported iPhone.

  • 1.14 GB. It downloads once inside the app from Private Diffusion's own server and picks up where it left off after an interruption. On iPhone and iPad the download keeps going with the app closed; on a Mac, quitting the app stops it, and it resumes at the next launch. You can delete it to reclaim the space and fetch it again later.

  • Tags or plain sentences both work. Start with the subject, then the setting and the light, and describe a drawing or a painting rather than a photo, because it learned from illustrations. Private Diffusion adds the quality tags "newest, masterpiece" to the front of every prompt for you, and a weight such as (fennec ears:1.3) strengthens that part of the prompt. Keep the prompt under about 1,000 characters: it reads the first 256 tokens and drops the rest.

    Build a prompt for Nanosaur 2 part by part

  • MIT License: Covers the image model and its VAE, not the pictures you make with them. Nothing in it asks anything of your images. Nanosaur 2's text encoder is Google's Gemma 3 270M, modified by Numen, so the Gemma Terms of Use applies as well; the link is on this page.

  • Yes. After the one-time download, Nanosaur 2 works without a connection. It turns your prompt into a finished image on the device, and you never sign in.

  • It is faster, the pictures hold up better, and nothing leaves your device. Private Diffusion ships Nanosaur 2 quantization aware distilled (QAD), trained against the full-size original, so it stays near-lossless at a fraction of the size. It runs on Apple's own framework: Core ML, across the Neural Engine and the GPU. ComfyUI runs models through PyTorch's MPS backend, the slowest inference framework on Apple platforms. ComfyUI's desktop app sends usage telemetry to PostHog and error reports to Datadog, with telemetry checked by default on first launch; Private Diffusion collects no data. And Private Diffusion is a sandboxed native Swift app with the hardened runtime on, where ComfyUI's desktop app is an Electron app that is not sandboxed.

  • Private Diffusion ships Nanosaur 2 quantization aware distilled (QAD), trained against the full-size original, so it stays near-lossless; Draw Things shrinks its models after training. It is faster, running on Core ML across the Neural Engine and the GPU. Private Diffusion collects no data, while Draw Things' App Store privacy label lists Purchases, Identifiers, and User Content. And it is a native Swift app built to Apple's Human Interface Guidelines, with Liquid Glass on macOS 26.4 or later, not a Mac Catalyst port.