base.toml

# General Configuration
[general]
# The anime name used for downloading animes from nyaa.si, and potentially for downloading images from Danbooru, unless --anime_name_booru is given
anime_name = "my_favorite_anime"
# Directory containing source files
src_dir = "path/to/src_dir"
# Directory to save output files
dst_dir = "path/to/dst_dir"
# Extra path component to add between dst_dir/[training|intermediate] and image_type
# The main difference between extra_path_component and image_type is that the
# latter can also be used in captions and considered as embeddings for pivotal tuning
extra_path_component = ""
# Stage number or alias to start from
start_stage = 1
# Stage number or alias to end at
end_stage = 7
# Directory to save logs. Set to None or none to disable.
log_dir = "logs"
# Prefix for log files, defaults to --anime_name if provided otherwise 'logfile'
log_prefix = {}
# Overwrite path in metadata. It has effect at stage 2 and 4-7
# Using by starting at stage 4 will prevent character information from being written to the original images
# Should never be used in general
overwrite_path = false
# Pipeline type that is used to construct dataset
remove_intermediate = false

# Metadata Loading and Saving
[metadata_handling]
# Extension of the grabber information files to load. Attributes from this file would overwrite those loaded from --load_aux.
# load_grabber_ext = ".tags"
load_grabber_ext = {}
# List of auxiliary attributes to load
# load_aux = ["processed_tags", "characters"] 
load_aux = []
# List of auxiliary attributes to save
# save_aux = ["processed_tags", "characters"]
save_aux = []

# Duplicate Detection Configuration
[duplicate_detection]
# Do not remove similar images
no_remove_similar = false
# Model used for duplicate detection
detect_duplicate_model = "mobilenetv3_large_100"
# Batch size for embedding computation used in duplicate detection
detect_duplicate_batch_size = 16
# Cosine similarity threshold for image duplicate detection
similar_thresh = 0.95

# Character Cropping Configuration
[character_cropping]
# Minimum size for character cropping
min_crop_size = 320
# Crop only images with head for character identification
crop_with_head = true
# Crop only images with face for character identification
crop_with_face = false
# The detection model level being used. The 'n' model runs faster with smaller system overhead
detect_level = "n"
# Use 3 stage crop to get halfbody and head crops
# This is slow and should only be called once for a set of images. Possible to use either at stage 2 or 4.
# Set to {} or comment this line to disable
use_3stage_crop = {}

# Character Clustering/Classification Configuration
[character_classification]
# Directory containing reference character images
character_ref_dir = "path/to/reference/character/images"
# The number of additional reference images to add to each character from classification result"
n_add_to_ref_per_character = 0
# Whether to disable matching character labels for noise images
no_extract_from_noise = false
# Whether to disable final filtering for character consistency
no_filter_characters = false
# Whether to keep unnamed clusters when reference images are provided or when characters are available in metadata
keep_unnamed_clusters = true
# Cluster merge threshold in character clustering
cluster_merge_threshold = 0.85
# Minimum cluster samples in character clustering
cluster_min_samples = 5
# The relative threshold for determining whether images belong to the same cluster for noise extraction and filtering
same_threshold_rel = 0.6
# The absolute threshold for determining whether images belong to the same cluster for noise extraction and filtering
same_threshold_abs = 20

# Dataset Construction Configuration
[dataset_construction]
# Overwrite existing trigger word csv
overwrite_emb_init_info = false
# Do not include cropped images in dataset
no_cropped_in_dataset = false
# Do not include original images in dataset
no_original_in_dataset = false
# Do not perform image resizing
no_resize = false
# Max image size that shorter edge aligns to
max_size = 768
# Dataset image extension
image_save_ext = ".webp"
# Filter repeated images again here
filter_again = true

# Tagging Configuration
[tagging]
# Whether to overwrite existing tags
overwrite_tags = false
# Method used for tagging
tagging_method = "wd14_convnextv2"
# Threshold for tagging
tag_threshold = 0.35

# General Tag Processing Configuration
[tag_processing]
# Mode to sort the tags
sort_mode = "score"
# Whether to append dropped character tags to the caption
append_dropped_character_tags = false
# Max number of tags to include in caption
max_tag_number = 30
# Path to the file containing blacklisted tags
blacklist_tags_file = "configs/tag_filtering/blacklist_tags.txt"
# Path to the file containing overlap tag information
overlap_tags_file = "configs/tag_filtering/overlap_tags.json"
# Path to the file containing character tag information
character_tags_file = "configs/tag_filtering/character_tags.json"
# Process tags from original tags instead of processed tags
process_from_original_tags = true
# Different ways to prune tags
prune_mode = "character_core"

# Core Tag Processing Configuration
[core_tag_processing]
# Number of directory levels to go up from the captioned directory when computing core tags
# Set to 1 here so that we compute a single time for all the image types
compute_core_tag_up_levels = 1
# Minimum frequency for a tag to be considered a core tag
core_frequency_thresh = 0.3
# Use existing core tag json instead of recomputing them
use_existing_core_tag_file = false
# The difficulty level up to which tags should be dropped
drop_difficulty = 2
# Whether to drop all core tags or not
drop_all_core = false
# The difficulty level from which tags should be used for embedding initialization
emb_min_difficulty = 1
# The difficulty level up to which tags should be used for embedding initialization
emb_max_difficulty = 2
# Whether to use all core tags for embedding initialization
emb_init_all_core = false
# Whether to append dropped character tags to wildcard
append_dropped_character_tags_wildcard = false

# Captioning Separation Configuration
[captioning_separation]
# For separating items of a single field of caption
caption_inner_sep = ", "
# For separating different fields of caption
caption_outer_sep = ", "
# For separating characters
character_sep = ", "
# For separating items of a single field of character
character_inner_sep = " "
# For separating different fields of character
character_outer_sep = ", "
# Separator for keep tokens for Kohya trainer
keep_tokens_sep = {}
# Where to put --keep_tokens_sep before
keep_tokens_before = "tags"

# Captioning Components Configuration
[captioning_components]
# Order of captions
caption_ordering = ["npeople", "character", "copyright", "image_type", "artist", "rating", "crop_info", "tags"]
# Probability to include number of people in captions
use_npeople_prob = 0
# Probability to include character info in captions
use_character_prob = 1
# Probability to include copyright info in captions
use_copyright_prob = 0
# Probability to include image type info in captions
use_image_type_prob = 1
# Probability to include artist info in captions
use_artist_prob = 0
# Probability to include rating info in captions
use_rating_prob = 1
# Probability to include crop info in captions
use_crop_info_prob = 1
# Probability to include tag info in captions
use_tags_prob = 1

# Folder Organization Configuration
[folder_organization]
# Number of directory levels to go up from the captioned directory when setting the source directory for the rearrange stage
rearrange_up_levels = 0
# Description of the concept balancing directory hierarchy
arrange_format = "n_characters/character"
# If have more than X characters put X+
max_character_number = 6
# Put others instead of character name if number of images of the character combination is smaller than this number
min_images_per_combination = 10

# Balancing Configuration
[balancing]
# Number of directory levels to go up from the rearranged directory when setting the source directory for the compute multiply stage
compute_multiply_up_levels = 1
# Minimum multiply of each image
min_multiply = 1
# Maximum multiply of each image
max_multiply = 100
# If provided use the provided csv to modify weights
weight_csv = "configs/csv_examples/default_weighting.csv"