Dataset Organization

After the entire process, you will get the following structure in /path/to/dataset_dir if you use the default configuration files and run booru and screenshots pipelines in parallel.

.
├── intermediate
│   ├── booru
│   │   ├── classified
│   │   ├── cropped
│   │   └── raw
│   └── screenshots
│       ├── animes
│       ├── classified
│       ├── cropped
│       └── raw
└── training
    ├── booru
    │   ├── 1_character
    │   ├── 2+_characters
    │   └── emb_init.json
    ├── screenshots
    │   ├── 0_characters
    │   ├── 1_character
    │   ├── 2_characters
    │   ├── 3+_characters
    │   └── emb_init.json
    ├── core_tag.json
    ├── emb_init.json
    └── wildcard.txt

💡 If --remove_intermediate is specified the folders classified and cropped are removed during the process.

The folder that should be used for training is /path/to/dataset_dir/training. Besides the training data, tt contains two important files.

• emb_init.json provides information for embedding initialization to be used for pivotal tuning (emb_init.json in the subfolders can be ignored).
• wildcard.txt provide the wildcard to be used with sd-dynamic-prompts.

You can put other folders, such as your regularization images in the training folder before launching the process so that they will be taken into account as well when we compute the repeat to balance the concept at the end.

Organization per Image Type

Each folder /path/to/dataset_dir/training/{image_type} is organized in the following way if --arrange_format is set to n_characters/character (the default value).

Level 1

├── ./0_characters
├── ./1_character
├── ./2_characters
├── ./3_characters
├── ./4+_characters

💡 Use --max_character_number n so that images containing more than n characters are all put together. If you don't want them to be included in the dataset. You can remove it manually.

Level 2

├── ./1_character
│   ├── ./1_character/AobaKokona
│   ├── ./1_character/AobaMai
│   ├── ./1_character/KuraueHinata
│   ├── ./1_character/KuraueHinata Hairdown
│   ├── ./1_character/KuraueKenichi
│   ├── ./1_character/KuraueMai
│   ├── ./1_character/KurosakiHonoka
│   ├── ./1_character/KurosakiTaiki
...

💡 Use --min_images_per_combination m so that character combinations with fewer than m images are all put in the folder character_others.
TODO: Add add an argument to optionally remove them.

The hierarchical organization allows to auto-balance between different concepts without too much need of worrying about the number of images in each class.

Multi-Anime Dataset and the Like

You can pass the argument --extra_path_component to replace {image_type} with {extra_path_component}/{image_type} in the aforementioned paths. This allows you for example to have a good organization when processing multiple animes in parallel.

Note that you will need to set --compute_core_tag_up_levels to 2 (or even higher number if --extra_path_component contains path separators) if you want to have a single wildcard and embedding initialization file for the entire dataset. Similarly, you may want to increase --rearrange_up_levels or --compute_multiply_up_levels to make sure that dataset balancing is computed from the root training folder.