LoRA_Easy_Training_Scripts in Docker

• https://github.com/derrian-distro/LoRA_Easy_Training_Scripts
• https://economylife.net/kohya-lora-install-use/

Environments

• Ubuntu 20.04 or later
• Docker Engine 23.0 or later
• NVIDIA Container Toolkit

Usage

1. Build Docker image

docker build . -t aoirint/lora_ets

2. Prepare Base Models

mkdir base_model
cd base_model

Stable Diffusion v2.x Based Models

Waifu Diffusion v1.4 Epoch 2

• https://huggingface.co/hakurei/waifu-diffusion-v1-4

wget 'https://huggingface.co/hakurei/waifu-diffusion-v1-4/resolve/main/wd-1-4-anime_e2.ckpt'
echo 'c76e0962bc60ccdc18e0dce387635b472b5a19038d637216030acbbe6eda2713  wd-1-4-anime_e2.ckpt' | sha256sum -c -

wget 'https://huggingface.co/hakurei/waifu-diffusion-v1-4/resolve/main/vae/kl-f8-anime2.ckpt'
echo 'df3c506e51b7ee1d7b5a6a2bb7142d47d488743c96aa778afb0f53a2cdc2d38d  kl-f8-anime2.ckpt' | sha256sum -c -

Waifu Diffusion v1.5 Beta 2

• https://huggingface.co/waifu-diffusion/wd-1-5-beta2

wget 'https://huggingface.co/waifu-diffusion/wd-1-5-beta2/resolve/main/checkpoints/wd-1-5-beta2-fp32.safetensors'
echo '764f93581d80b46011039bb388e899f17f7869fce7e7928b060e9a5574bd8f84  wd-1-5-beta2-fp32.safetensors' | sha256sum -c -

Stable Diffusion v1.x Based Models

Waifu Diffusion v1.3

• https://huggingface.co/hakurei/waifu-diffusion-v1-3

wget 'https://huggingface.co/hakurei/waifu-diffusion-v1-3/resolve/main/wd-v1-3-full-opt.ckpt'
echo '10912b9a6d773ea7c299c0563d10538ada04ade81837362b6c0c67be4df937c1  wd-v1-3-full-opt.ckpt' | sha256sum -c -

Anything v5

• https://civitai.com/models/9409?modelVersionId=29588

echo '7f96a1a9ca9b3a3242a9ae95d19284f0d2da8d5282b42d2d974398bf7663a252  anything-v5-prt-re.safetensors' | sha256sum -c -

Anything v4.5

• https://huggingface.co/andite/anything-v4.0

wget 'https://huggingface.co/andite/anything-v4.0/resolve/main/anything-v4.5.ckpt'
echo 'fbcf965a62d9d82e935d3d17e97522c29f44550aa9e120a6886f19b578521ec5  anything-v4.5.ckpt' | sha256sum -c -

Anything v4.0

• https://huggingface.co/andite/anything-v4.0

wget 'https://huggingface.co/andite/anything-v4.0/resolve/main/anything-v4.0.ckpt'
echo '3b26c9c497c923a07ab8d55f2921cf44749535e4b0c890c5c37968e4c90e7258  anything-v4.0.ckpt' | sha256sum -c -

wget 'https://huggingface.co/andite/anything-v4.0/resolve/main/anything-v4.0.vae.pt'
echo 'f921fb3f29891d2a77a6571e56b8b5052420d2884129517a333c60b1b4816cdf  anything-v4.0.vae.pt' | sha256sum -c -

Anything v3

• https://civitai.com/models/9409?modelVersionId=11162
• https://huggingface.co/Linaqruf/anything-v3.0

echo '8712e20a5d65b6acaa743e8a74961eadfdf846a2c9a32160d80a80cba13ad475  anything-v3.ckpt' | sha256sum -c -

echo 'f921fb3f29891d2a77a6571e56b8b5052420d2884129517a333c60b1b4816cdf  anything-v3.vae.pt' | sha256sum -c -

# VAE embbeded
echo 'abcaf14e5acb8023c79c3901f8ffc04eb3c646d7793f3b36a439bf09e32868cd  anything-v3-full.safetensors' | sha256sum -c -

ACertainThing

• https://huggingface.co/JosephusCheung/ACertainThing

wget 'https://huggingface.co/JosephusCheung/ACertainThing/resolve/main/ACertainThing.ckpt'
echo '866946217b513157b12ff9b1eae2279e98ae34ece178e34eac536f2e831c101c  ACertainThing.ckpt' | sha256sum -c -

ACertainty

• https://huggingface.co/JosephusCheung/ACertainty

wget 'https://huggingface.co/JosephusCheung/ACertainty/resolve/main/ACertainty.ckpt'
echo 'a64573359af0f1071ef01d0dc93df2bc90eb1d0bcf3e26058fbf5aeff37c6462  ACertainty.ckpt' | sha256sum -c -

2. Prepare Training Images and Prompts

Create work/20230225_001/img/20_my girl directory. Place images and their prompt text files.

On WebUI, prompt text files can be generated by the extension toriato/stable-diffusion-webui-wd14-tagger.

• work/20230225_001/img/20_my girl
  • 0001.png
  • 0001.txt
  • 0002.jpg
  • 0002.txt
  • ...

3. Prepare Regularization Images

Create work/20230225_001/reg_img/1_girl directory. Place transparent PNG images (the number of them should be the same as the number of the training images). It is said to improve training quality.

On WebUI, transparent PNG images can be generated by the extension hunyaramoke/Generate-TransparentIMG.

Or you can use this.

• work/20230225_001/reg_img/1_girl
  • transparent_1.png
  • transparent_2.png
  • ...

4. Prepare JSON Config

• https://github.com/derrian-distro/LoRA_Easy_Training_Scripts#list-of-arguments

Create work/20230225_001/config.json.

Stable Diffusion v2.x Based Models

{
  "base_model": "/base_model/wd-1-4-anime_e2.ckpt",
  "v2": true,
  "v_parameterization": true,
  "clip_skip": null,
  "img_folder": "/work/20230225_001/img",
  "reg_img_folder": "/work/20230225_001/reg_img",
  "output_folder": "/work/20230225_001/output",
  "num_epochs": 10,
  "batch_size": 1,
  "save_every_n_epochs": 1
}

Stable Diffusion v1.x Based Models

{
  "base_model": "/base_model/ACertainThing.ckpt",
  "img_folder": "/work/20230225_001/img",
  "reg_img_folder": "/work/20230225_001/reg_img",
  "output_folder": "/work/20230225_001/output",
  "num_epochs": 10,
  "batch_size": 1,
  "save_every_n_epochs": 1
}

4. Prepare Output and Cache directory

Create work/20230225_001/output directory.

Create cache/huggingface/hub directory.

5. Run Training

Directory structure

• base_model/
• work/20230225_001/
  • img/20_my girl/
    • 0001.png
    • 0001.txt
    • 0002.jpg
    • 0002.txt
    • ...
  • reg_img/1_girl/
    • transparent_1.png
    • transparent_2.png
    • ...
  • output/
• cache/huggingface/hub/

Command

chown -R 1000:1000 base_model/ work/ cache/

docker run --rm \
  --gpus all \
  -v "./base_model:/base_model" \
  -v "./work:/work" \
  -v "./cache/huggingface/hub:/home/user/.cache/huggingface/hub" \
  aoirint/lora_ets \
  --load_json_path=/work/20230225_001/config.json