Minimal Apple Silicon fork of ostris/ai-toolkit.

This fork exists for a very specific workflow:

• train locally on macOS with device: mps
• generate sample previews on another machine through Draw Things
• allow the remote preview model to be different from the local training model

This README only covers macOS on Apple Silicon.

• Local training is meant to run on Apple Silicon with MPS.
• Preview sampling is remote, never local in-process generation.
• Preview sampling is configured through sample.drawthings.
• The remote sample model can be different from the model you are training on your Mac.
• Returned previews are saved by AI Toolkit, not into the Draw Things gallery on the remote machine.

• Apple Silicon Mac
• macOS
• Xcode Command Line Tools
• Homebrew
• Python 3.12
• Node 22
• Git

xcode-select --install

brew install [email protected] node@22 git
echo 'export PATH="/opt/homebrew/opt/[email protected]/bin:/opt/homebrew/opt/node@22/bin:$PATH"' >> ~/.zprofile
source ~/.zprofile

git clone https://github.com/hexvalid/ai-toolkit-hex.git
cd ai-toolkit-hex

python3.12 -m venv venv
source venv/bin/activate

pip install --upgrade pip
pip install --no-cache-dir torch==2.9.1 torchvision==0.24.1 torchaudio==2.9.1
pip install -r requirements.txt

cd ui
npm run build_and_start

Open http://localhost:8675.

Training happens on your Mac. Sample previews happen on another machine running Draw Things remote access.

In the job config or UI, set:

• device: mps
• sample.drawthings.server
• sample.drawthings.port
• sample.drawthings.model
• sample.drawthings.shared_secret if your remote server requires it

The remote Draw Things model can be different from model.name_or_path.

source venv/bin/activate
python run.py path/to/config.yml

• You still need the correct Hugging Face access/token for gated models.
• This fork is intentionally optimized around the MPS + remote-preview workflow, not general multi-platform setup docs.

This project is based on ostris/ai-toolkit. Credit for the core project, training system, and the vast majority of the original work goes to the upstream project and its contributors.