Mochi Diffusion

Overview

Mochi Diffusion is a native macOS application for running Stable Diffusion and FLUX.2 Klein models locally. It uses Apple's Core ML Stable Diffusion implementation for performance and reduced memory requirements, plus the Iris pipeline for FLUX.2 Klein models with native Metal GPU acceleration.

Key Features

• Generates images locally and completely offline.
• Native macOS app built with SwiftUI.
• Uses Apple's Core ML Stable Diffusion implementation for speed and reduced memory requirements, especially on Apple Silicon Macs with Neural Engine.
• Runs FLUX.2 Klein models through the Iris pipeline with native Metal GPU acceleration.
• Saves generated images as PNG with generation settings readable by AUTOMATIC1111-compatible tools and Civitai uploads.
• Includes a built-in gallery with import, save, and sync support.

Use Cases

• Local image generation on Apple Silicon Macs.
• Offline image generation workflows where data should remain on the device.
• Stable Diffusion generation using a starting image or ControlNet.
• FLUX.2 Klein generation using up to four input images.

Getting Started

• Download the latest Mochi Diffusion release.
• Download or convert Core ML Stable Diffusion models, or download FLUX.2 Klein model components.
• Place models in the MochiDiffusion model folder under the home directory; the location can be customized in Settings.
• For Core ML Stable Diffusion, create a model folder named as it should appear in the app and move or extract the converted model there.
• For FLUX.2 Klein, place text_encoder, tokenizer, transformer, and vae components in the model folder.

Deployment & Requirements

• Apple Silicon (M1 and later).
• macOS 15.6 and later.
• Xcode 26.0 or later to build from source.
• Core ML Stable Diffusion or FLUX.2 Klein models are required to use Mochi Diffusion.

Before You Adopt

• License: GPL-3.0. Review its terms before using, modifying, or distributing the project.
• Mochi Diffusion requires separately obtained Core ML Stable Diffusion or FLUX.2 Klein models to generate images.
• The app does not collect data or telemetry, and generation happens locally without sending anything to the cloud.
• First use of a model may take up to two minutes for Neural Engine compilation, after which generations are faster.
• FLUX.2 Klein may need constrained input image dimensions due to pipeline attention-budget limits; Mochi resizes inputs and shows the new sizes.