sd-scripts-docker

Dockerfile for kohya-ss/sd-scripts.

Requirements

• Ubuntu 24.04 or later
• Docker Engine 29 or later
• NVIDIA Container Toolkit
• NVIDIA GeForce RTX 4000 series, 5000 series
  • 1000 series does not work due to CUDA compatibility.
  • 2000 series and 3000 series might work, but untested.

Usage

• Replace accelerate launch with sudo docker run --rm --gpus all aoirint/sd_scripts.
• Training command will run in the container by a general user (UID=1000).

WD14 Captioning (ONNX)

mkdir -p "./cache/wd14_tagger_model_cache"
sudo chown -R 1000:1000 "./cache/wd14_tagger_model_cache"

# If your cache is broken, execute
# rm -rf ./cache/wd14_tagger_model_cache/wd14_tagger_model

sudo docker run --rm --gpus all \
  -v "./work:/work" \
  -v "./cache/wd14_tagger_model_cache:/wd14_tagger_model_cache" \
  aoirint/sd_scripts \
  finetune/tag_images_by_wd14_tagger.py \
  --model_dir "/wd14_tagger_model_cache/wd14_tagger_model" \
  --onnx \
  /work/my_dataset-20230715.1/img

Training LoRA-LierLa U-Net only with DreamBooth、キャプション方式 for Animagine XL 4.0 Zero (Stable Diffusion XL)

Create permanent directories to mount on container.

mkdir -p "./base_model" "./work" "./cache/huggingface/hub"
sudo chown -R 1000:1000 "./base_model" "./work" "./cache/huggingface/hub"

Download animagineXL40_v4Zero.safetensors from Animagine XL 4.0 Zero.

wget -O "animagineXL40_v4Zero.safetensors" "https://civitai.com/api/download/models/1409042?type=Model&format=SafeTensor&size=full&fp=fp16"
echo "f15812e65c2ea7f4e19ce37fb2a8445eb65c64da450a508dd9c8f237c73f6bb8  animagineXL40_v4Zero" | sha256sum -c -

Prepare a dataset directory work/my_dataset-20230715.1 and a config file work/my_dataset-20230715.1/config.toml following train_README.

Set file ownership UID:GID = 1000:1000 (sudo chown -R 1000:1000 "./work").

You can also choose another directory structure to modify config.toml and the training command.

• work/my_dataset-20230715.1/
  • config.toml
  • img/
    • 0001.png
    • 0001.txt
    • 0002.png
    • 0002.txt
    • ...
  • reg_img/
    • transparent_1.png
    • transparent_2.png
    • ...
  • sample_prompts.txt
  • output/
  • logs/

This is an example sample_prompts.txt. Use the output of WD14 tagger as reference. Add new lines for multiple samples.

shs 1girl, 1girl, solo, simple background, white background, masterpiece, high score, great score, absurdres --w 1024 --h 1024 --d 42 --s 28 --l 5 --ss euler_a --n lowres, bad anatomy, bad hands, text, error, missing finger, extra digits, fewer digits, cropped, worst quality, low quality, low score, bad score, average score, signature, watermark, username, blurry

This is an example config.toml.

[general]
enable_bucket = true

[[datasets]]
resolution = 1024
batch_size = 1

  [[datasets.subsets]]
  image_dir = '/work/my_dataset-20230715.1/img'
  caption_extension = '.txt'
  num_repeats = 20

  [[datasets.subsets]]
  is_reg = true
  image_dir = '/work/my_dataset-20230715.1/reg_img'
  class_tokens = '1girl'
  num_repeats = 1

Execute training.

sudo docker run \
  --rm \
  --gpus all \
  -v "./base_model:/base_model" \
  -v "./work:/work" \
  -v "./cache/huggingface/hub:/home/user/.cache/huggingface/hub" \
  aoirint/sd_scripts \
  --num_cpu_threads_per_process=1 \
  sdxl_train_network.py \
  --seed=42 \
  --pretrained_model_name_or_path="/base_model/animagineXL40_v4Zero.safetensors" \
  --dataset_config="/work/my_dataset-20230715.1/config.toml" \
  --output_dir="/work/my_dataset-20230715.1/output" \
  --output_name="my_dataset-20230715.1" \
  --save_model_as="safetensors" \
  --logging_dir="/work/my_dataset-20230715.1/logs" \
  --prior_loss_weight=1.0 \
  --max_train_epochs=5 \
  --learning_rate=1e-4 \
  --optimizer_type="AdaFactor" \
  --xformers \
  --mixed_precision="fp16" \
  --cache_latents \
  --cache_text_encoder_outputs  \
  --gradient_checkpointing \
  --save_every_n_epochs=1 \
  --sample_at_first \
  --sample_every_n_epochs=1 \
  --sample_prompts="/work/my_dataset-20230715.1/sample_prompts.txt" \
  --sample_sampler="euler_a" \
  --network_module="networks.lora" \
  --network_train_unet_only

Maintenance

• Updating sd-scripts

Release procedure

Releases are driven by the root VERSION file and the Git tags on GitHub. To update the bundled sd-scripts version before a release, follow Updating sd-scripts first.

1. Update VERSION to the version to publish.
   • Stable releases use SemVer without a prerelease suffix, such as 0.1.0.
   • Prereleases use SemVer with a prerelease suffix, such as 0.1.0-rc.1.
2. Commit the VERSION change and merge it to main.
3. The build workflow checks whether v<VERSION> already exists on GitHub.
   • If the tag does not exist and VERSION is stable, it creates a latest GitHub Release and publishes Docker images tagged v<VERSION> and latest.
   • If the tag does not exist and VERSION is a prerelease, it creates a prerelease GitHub Release and publishes the Docker image tagged edge.
   • If VERSION is 0.0.0, it is treated as an edge build and only the edge Docker image is updated.
   • If the tag already exists, the push is treated as an edge build and only the edge Docker image is updated.

License

This repository's Dockerfile, documentation, and project-specific files are licensed under the MIT License. See LICENSE.

The published Docker image bundles kohya-ss/sd-scripts at the commit specified by SD_SCRIPTS_VERSION in the Dockerfile. sd-scripts is primarily licensed under the Apache License 2.0, with some portions under separate license terms. See THIRD_PARTY_NOTICES.md and the upstream license information for details.