Wenwu2000/MobileDiffuser-SD3-medium

MobileDiffuser SD3 Medium Core ML Models

This repository contains the Core ML model bundle used by MobileDiffuser for on-device Stable Diffusion 3 Medium inference on iPhone.

Contents

• coremlsd3_2step/: SD3 Medium two-step Core ML resources.
• coremlsd3_4step/: SD3 Medium four-step Core ML resources.
• checkpoints/sd3-medium-2step.safetensors: source two-step distilled checkpoint used to build the two-step Core ML bundle.
• checkpoints/sd3-medium-4step.safetensors: source four-step distilled checkpoint used to build the four-step Core ML bundle.

Each bundle is organized for MobileDiffuser's split MMDiT runtime:

• TextEncoder.mlmodelc
• TextEncoder2.mlmodelc
• MultiModalDiffusionTransformerConditioning.mlmodelc
• MultiModalDiffusionTransformerStage*.mlmodelc
• VAEDecoder.mlmodelc

The MMDiT model is split into multiple ANE-friendly stages to reduce live activation pressure and avoid out-of-memory termination on iPhone 15 Pro class devices. The app loads the selected resource directory and runs either the two-step or four-step scheduler configuration.

Usage With MobileDiffuser

Clone this repository next to the MobileDiffuser app repository, or copy the two resource directories into the MobileDiffuser project root:

cp -R coremlsd3_2step /path/to/MobileDiffuser/
cp -R coremlsd3_4step /path/to/MobileDiffuser/

Then open MobileDiffuser.xcodeproj, make sure both folders are included in the app target resources, configure your signing team, and deploy to device.

Git LFS

The model weights are stored with Git LFS. After cloning, run:

git lfs install
git lfs pull

If the .mlmodelc directories contain small pointer files instead of real model weights, LFS objects were not pulled successfully.

The .safetensors checkpoints are also stored through LFS. They are provided so the Core ML bundles can be reproduced or re-converted with different splitting or quantization settings.

Notes

• These resources are intended for 512x512 generation.
• The runtime is optimized for cpuAndNeuralEngine.
• The source training checkpoint is not included here.
• Make sure your use of the original SD3 Medium weights complies with the upstream model license.