Kohaku XL Gamma
A SDXL anime base model aims to create unique artworks.
Not designed for reproducing copyright characters or artist style
Introduction
This model can be seen as a derivative of Animagine XL 3.0 project.
Basically I'm collaborating with Linaqruf for making better Anime base model (and it is obvious that we have different goal/target)
We share our models and technique to improve our models' quality.
And that is also how this model been created.
Models
Base7
Kohaku-XL base7 is resumed from beta7 and use same dataset that beta series have used. But this time I use my own metadata system to create captions. (Can be taken as advanced version of what linaqruf used, will open source it soon)
The metadata database can be downloaded here:
KBlueLeaf/danbooru2023-sqlite · Datasets at Hugging Face
Trainin details:
LR: 8e-6/2e-6
Scheduler: constant with warmup
Batch size: 128 (batch size 4 * grad acc 16 * gpu count 2)
Base 8
Kohaku-XL base8 is as same as base7 but use 5e-6/1e-6 LR, and it have a full epoch on 1.48M images.
Gamma rev1
Kohaku-XL Gamma rev1 is a merged model which combine the learned diff from anxl3 and kohaku xl base 7. With this forumla:
gamma rev1 = beta7 + 0.8 * (anxl3 - anxl2) + 0.5 * (base7 - beta7)
Gamma rev2
Kohaku-XL gamma rev2 is merged with this equation:
gamma rev2 = beta7 + 1.0 * (anxl3 - anxl2) + 0.25 * (base8 - beta7)
I also use a MBW recipe to merge with Anxl3 directly:
0,0.1,0.1,0,0.1,0.1,0,0.1,0.1,0,0,0,0,0,0.05,0.05,0.05,0.05,0.05,0.05
Why merge:
First of all, (anxl3 - anxl2) can be take as "What Anxl3 have learned from the training resumed from anxl2". And since the "learned things" are basically "meta tags" and some character/style information, it is good to merge the diff from anxl3/anxl2.
Or you can take this merge as "merged with 2 lora/lycoris model, one is extracted from anxl3 and another is extracted from base7/8"
The very low weight MBW merge is to fix some over-trained artifact in the merged model.