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.

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)

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)

Usage

Parameter

This model is trained under 768×1024 to 1024×1024 ARB. It is recommended to use pixel count within 786432 ~ 1310720.
Recommended CFG scale is 4~7.
Sampler should not matter.

Tagging

This model use my own system for quality tags or something like that.
So although this model combine the diff weight from anxl3, I will still recommend user to use mine (or both) tagging system.

This model is trained with artists’ name as a tag so you can use artists tag as “stylize tag”. But since my training config is not designed for learning style. I don’t think you can actually reproduce the style of any specific artists. Just use artists tag as style refiner tag.
For character tags, as same as artists. The model can accept character tags but I won’t ensure its ability on reproducing any character.

The format of prompt is as same as anxl3. (You can check the sample images I post)

Rating tags:

• General: safe

• Sensitive: sensitive

• Questionable: nsfw

• Explicit: explicit, nsfw

Quality tags (Better to worse):

• Masterpiece

• best quality

• great quality

• good quality

• normal quality

• low quality

• worst quality

Year tags (New to Old):

• newest

• recent

• mid

• early

• old

You may meet some subtle mosaic-like artifact, that may be caused by high-lr or bad resizing/image encoding.
I will try to fix it in next version. For now, try to use R-ESRGAN anime6b or SCUNet models for fixing it.

Future plan

Since my dataset have some resize/webp artifacts that will harm the models. I will recreate my dataset based on my new system (and opensource it once I done it).

The next plan is to train model on larger (3M~6M) dataset with better configuration (which will require A100s and I plan to spend about 2000~10000 USD on it, if you like my works, consider to sponsor me via buy-me-a-coffee or some BTC-sutff, you can get the link from my github profile: KohakuBlueleaf (Kohaku-Blueleaf) (github.com))

Appendix

When I try to generate some sample images, I found my merge method actually goes wrong at first (due to some bugs in LyCORIS’ utils).
Although the final version of rev1 is not a bugged one, some of sample images I post is actually generated by it.
I will upload the bugged one and base7 onto my Huggingface:
KBlueLeaf (Shih-Ying Yeh) (huggingface.co)

If you want to follow up the progress of next version (or some other projects)
Checkout my homepage:
Kohaku’s Homepage (kblueleaf.net)

Licence

Important

This model is licensed under faipl-1.0-sd just like anxl3:
Freedom of Development (freedevproject.org)