LEAF: Latent Diffusion with Efficient Encoder Distillation for Aligned Features in Medical Image Segmentation

Performance

Setup

1. Clone the repository:

git clone https://github.com/lispear/LEAF.git
cd LEAF-master

2. Install dependencies (requires conda):

conda create -n leaf python=3.11.11 -y
conda activate leaf
pip install -r requirements.txt 

Training

1. Create assets directory:

mkdir assets
cd assets

2. Prepare pre-trained models:

• Download U-Net and VAE
• Place the downloaded files in the assets folder.
• Extract the weights:

unzip kl-f8.zip -d vae
unzip lsun_churches.zip -d unet
cd ..
python extract_weights.py

3. Generate training configuration:

python src/util/config_util.py --output config.yaml

4. Run training script:

accelerate launch \
    --num_processes 1 \
    --num_machines 1 \
    --mixed_precision 'no' \
    --dynamo_backend 'no' \
    train.py --config config.yaml