Just a repo for sharing random CoreML & Stable Diffusion files & info.

Sept 26,2023:

I just created a fresh miniconda environment for model conversions.

It is working for SD-1.5 type models, including the ControlledUnet, both original and split-einsum-v1.

It seems faster than the environments that I've run before. Converting ControlNets also works.

I have not tried it with split-einsum-v2, or SDXL.

It uses mostly the current versions of available packages, specifically:

macOS 14.0
python==3.9
coremltools==7.0
diffusers==0.21.2
python-coreml-stable-diffusion (pip -e . from a current git clone of ml-stable-diffusion)
safetensors==0.3.3
scikit-learn==1.1.2
torch==2.0.0
torch2coreml.py from the current diffusers main repo

The converted models work in both the standard Mochi Diffusion 4.2 and a version I built with Xcode 15 for macOS 14 using a git clone of the current Mochi Diffusion main branch.