openjug anime v1.0 post

(civitai interface is a bit wierd. Guess I'll repeat myself a bit here)

This is intended to be a 100% "open" finetune. I tell you EVERYTHING I used to make it!

The base is Juggernaut9.

Training settings

I trained it with OneTrainer, using the following settings:

(BF16) LION, constant rate, batch=1 warmup=0
Learning Rate=1e-05, epochs=100
Attention=XFORMERS, EMA update step=100,, EMA=cpu, gradient accumilation=on
autocast cache=0, resolution=1024
NO UNET TRAINING

Dataset

The dataset was given as a single "concept" definition in OneTrainer, with tags specified per-file.

The entire 605 image dataset can be downloaded via the scripts in

https://huggingface.co/datasets/ppbrown/danbooru-cleaned/tree/main/SUBSETS

using the input list

https://huggingface.co/datasets/ppbrown/danbooru-cleaned/blob/main/SUBSETS/files.onegirlscene605.txt

(The scripts should grab both the .jpg files, AND the matching .txt files)

The burden

Casual visitors will have no idea just how much time this took.

From narrowing the dataset down from 50,000 danbooru images, to trying multiple base models, multiple settings.. over, and over.... ugh.

I hope the results are a nice thing for some people. I also hope that being open with the pieces of it

This was trained using a 4090. The FINAL training of 60,000 steps didnt take that long. 6-10 hours maybe?

It was all the experimental steps alone the way, that tooks days upon days of experimenting and tweaking....

I have no more words. They are all gone now :)