dataset-example.toml

# Lora training is basing on the method of DreamBooth+Caption(without using class_images to aviod over-fitting). Besides DreamBooth+Caption, Lora can also use other methods, such as DreamBooth+identifier+class and classic text2image fine tuning. Detailed info you can refer to https://github.com/kohya-ss/sd-scripts/blob/main/train_README-ja.md

[general]
enable_bucket = true

# DreamBooth caption based datasets
[[datasets]]
resolution = 512
batch_size = 2
shuffle_caption = true
caption_extension = '.txt'
keep_tokens = 0

  [[datasets.subsets]]
  image_dir = './images/sample1'
  num_repeats = 10

#  [[datasets.subsets]]
#  image_dir = './images/sample2'
#  num_repeats = 10

#  [[datasets.subsets]]
#  is_reg = true
#  image_dir = './images/reg'
#  class_tokens = 'girl'
#  num_repeats = 1  


# DreamBooth class+identifier based datasets
#[[datasets]]
#resolution = 512
#batch_size = 2

#  [[datasets.subsets]]
#  image_dir = './images/sample1'
#  class_tokens = 'wta girl'
#  num_repeats = 10

#  [[datasets.subsets]]
#  image_dir = './images/sample2'\
#  class_tokens = 'zyl girl'
#  num_repeats = 10

#  [[datasets.subsets]]
#  is_reg = true
#  image_dir = './images/reg'
#  class_tokens = 'girl'
#  num_repeats = 1  


# Classic fine tuning based dataset
#[[datasets]]
#resolution = [768, 768]
#batch_size = 2

#  [[datasets.subsets]]
#  image_dir = './images/sample3'
#  metadata_file = './images/sample3/xxx_md.json'