Skip to content

A training template for Low-rank adaptation diffusion models.

License

Notifications You must be signed in to change notification settings

3ee-Games/lora-concepts-template

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

3 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

lora concepts template

This is a template that can be used for training multiple concepts and regularization images for lora (Low-rank Adaptation) models.

Multiple concept loRA structure

Directory Description
image_dir contains all input images and captioning files, prefixed by the total number of epoch per image
reg_dir contains all regularization images per concept, prefixed by the total number of repeats per reg image
log records of each training log
output the final output binary (safesensor, pt, etc)
config_v1_example.json the config for the training session (rename to whatever you want)

Balancing datasets

Balance the datasets so that the concept folders indicate the number of times they should be repeated during training.

For example, let's say you have four concepts and each of those has 14 images per concept, except for one, which has 28.

To balance this for training steps per concept per epoch 5200, you would divide the repeats by the total amount of images per concept:

5200 / 14 = 371 5200 / 28 = 186

In the example, the name of the concept directories would be:

  • 186_conceptA
  • 371_conceptB
  • 371_conceptC
  • 371_conceptD

bmaltais utilities

The below utility will ensure that each concept folder in the dataset folder is used equally during the training process of the dreambooth machine learning model, regardless of the number of images in each folder:

https://github.com/bmaltais/kohya_ss/blob/cdf84e2f4b65517c396ffb72474616a57bc109b6/library/dataset_balancing_gui.py