Docker for Dreambooth Stable Diffusion

This is a Docker Setup for Dreambooth to train personalized stable diffusion models.

See wsl2-setup.md for infos about setting up WSL2 and Docker under Windows.

Install

• clone repo
• Build with: docker compose -f docker-compose.build.yaml build

One-Time Setup

• register on huggingface and create a new Access Token here: https://huggingface.co/settings/tokens
• copy the token and set it in a file named .env in the form TOKEN=yourtoken. (see also example.env)

Configuration

• make sure the folders checkpoint_output/, class_images and output/ are empty before training a new checkpoint.

  (if they don't exist, they are created on the first start)

• put training images into instance_images/ directory.

• edit app/train.sh file in a text editor.

  • only uncomment the type of training you want to use.
  • edit INSTANCE_PROMPT and CLASS_PROMPT

• Edit settings in accelerate-config.yaml or run docker compose -f docker-compose.build.yaml exec accelerate config --config_file=/app/examples/dreambooth/accelerate-config.yaml to configure training acceleration settings.

Training

• run container with docker compose -f docker-compose.build.yaml up -d
• run docker compose -f docker-compose.build.yaml exec dreambooth ./train.sh
• get the generated checkpoint file to use with most SD UIs (like https://github.com/AUTOMATIC1111/stable-diffusion-webui) from checkpoint_output/model.ckpt.

Helper

• open shell into container with: docker compose -f docker-compose.build.yaml exec dreambooth bash

Sources

• https://github.com/ShivamShrirao/diffusers/tree/main/examples/dreambooth
• https://gist.github.com/jachiam/8a5c0b607e38fcc585168b90c686eb05