Apple Silicon’s offline image foundry

What it does

Mochi Diffusion is a native macOS app that generates images from text prompts using Stable Diffusion or FLUX.2 Klein, entirely on your Mac. It wraps Apple’s Core ML pipeline and the iris.c FLUX implementation in a SwiftUI interface, keeping all computation local and offline. Think of it as a polished frontend for the Neural Engine with a built-in gallery and EXIF metadata tagging.

The interesting bit

The project does not reinvent the diffusion wheel; it repackages existing pipelines into a consumer Mac application that claims to use only ~150 MB of memory when running on the Neural Engine. Adding FLUX.2 Klein support via the Iris pipeline without model conversion shows an effort to support newer architectures within Apple’s hardware constraints.

Key highlights

• Runs Stable Diffusion and FLUX.2 Klein locally on Apple Silicon using Core ML and the Neural Engine.
• Claims ~150MB memory usage with Neural Engine; runs offline with no cloud dependency.
• Supports Image2Image, ControlNet, and custom Core ML models.
• Saves generation metadata (prompts) directly into image EXIF.
• Native SwiftUI app with a built-in gallery.

Caveats

• Apple Silicon only; requires macOS 15.6 or later.
• First-time model compilation can take up to two minutes before subsequent generations speed up.
• Core ML Stable Diffusion support is limited to converted split_einsum or original formats with differing compute-unit compatibility, and FLUX.2 Klein models must follow a specific directory hierarchy inside the app’s model folder.

Verdict

Worth a look if you want a native, offline image generator on a modern Mac and prefer not to wrestle with Python environments. Skip it if you are on Intel, need cross-platform tooling, or want a fully automated model setup.