Dive into an in-depth tutorial on using LyCORIS for style training in Stable Diffusion, focusing on LoCon and LoHa techniques as an extension of LoRA. Learn about required resources, LyCORIS overview, and detailed training settings for both LoCon and LoHa. Explore grid analyses comparing epochs and weights for different dimensions, and understand how to finalize hyperparameters. Discover the process of creating and managing style datasets using tools like BooruDatasetTagManager. Follow along as various grid analyses are performed for different artists' styles, and gain insights into adjusting LoHa hyperparameters for optimal results. Conclude with valuable closing thoughts on implementing these advanced techniques in your AI art creation process.
Overview
Syllabus
Intro
Required resources
LyCORIS overview
LoCon training settings
redjuice LoCon Epoch vs. Weight grid analysis 32 dims
redjuice LoCon Epoch vs. Weight grid analysis 40 dims
Vanilla LoRA
Explaining how I finalized my hyperparameters
LoHa training settings
redjuice LoHa grid analysis Conv. Dim = 4, Conv. Alpha = 2
redjuice LoHa grid analysis Conv. Dim = 4, Conv. Alpha = 1
swd3e2 style dataset overview
BooruDatasetTagManager
swd3e2 LoCon grid analysis 40 dims
swd3e2 LoHa grid analysis Conv. Dim = 4, Conv. Alpha = 1
Adjusting LoHa hyperparameters Part 1
swd3e2 LoHa grid analysis, Conv. Dim = 4, Conv. Alpha = 2, Adam8Wbit, fixed initial learning rate
Adjusting LoHa hyperparameters Part 2
swd3e2 LoHa grid analysis, Conv. Dim = 14, Conv. Alpha = 14, DAdaptation
Closing thoughts
Taught by
kasukanra
Related Courses
-
Stable Diffusion - Create a PNGtuber with Character LoRA
-
Stable Diffusion 3 2B Medium Training with Kohya and SimpleTuner - Full Finetune and LoRA
-
Stable Diffusion - How to Train LoRA for Style - LoRA Part 1
4.0 -
Stable Diffusion- Training SDXL 1.0 - Finetune, LoRA, D-Adaptation, Prodigy
-
Perfect Style Lora Training for SDXL - Amazing Results
-
LORA Training in Stable Diffusion - Less Than 7GB VRAM
