SCAL-SDT

Stable Diffusion trainer with scalable dataset size and hardware usage.

[!] IN EARLY DEVELOPMENT, CONFIGS AND ARGUMENTS SUBJECT TO BREAKING CHANGES

Features

• Can run with 10G or less VRAM without losing speed thanks to xformers memory efficient attention and int8 optimizers.
• Aspect Ratio Bucketing
• DreamBooth
• CLIP skip
• WandB logging

Getting Started

Install Requirements

Linux is recommended, on Windows you have to install bitsandbytes manually for int8 optimizers.

Conda

conda env create environment.yml
conda activate ssdt

Pip

CUDA toolkit and torch should be installed manually.

pip install -r requirements.txt

Config

Documentation: configs/README.md. (Link)

configs/native.yaml (for native training) and configs/dreambooth.yaml (for DreamBooth) provided as examples.

Run

If you are running native training, proceed to the next step.
If you are running DreamBooth, run this to generate class (regularization) images:

python gen_class_imgs.py --config configs/your_config.yaml

Then run the training:

python train.py --config configs/your_config.yaml

After Training

Note although the checkpoints have .ckpt extension, they are NOT directly usable to interfaces based on the official SD code base like WebUI. To convert them into SD checkpoints:

python convert_to_sd.py PATH_TO_THE_CKPT OUTPUTDIR --no-text-encoder --unet-dtype fp16

--no-text-encoder --unet-dtype fp16 results a ~2GB checkpoint, containing fp16 UNet and fp32 VAE weights, WebUI supports loading that. For further reducing checkpoint size to ~1.6GB if target clients have external VAE already, add --no-vae to remove VAE weights from checkpoint, leaving fp16 UNet weights only.

If you are not using WebUI and having issues, remove --no-text-encoder.

TPUs or other computing units?

You may change trainer.accelerator. (Docs)

Advanced

• Config Documentation
• SCAL-SDT Wiki