Type | |
Stats | 75 2 5 |
Reviews | (7) |
Published | Sep 25, 2024 |
Base Model | |
Training | Steps: 60,500 Epochs: 100 |
Trigger Words | anime |
Hash | AutoV2 BF466B51CA |
Welcome!
This is an experiment in OPEN finetuning. This means that I will give you everything I used to create this finetune.
Warning: do not use default CFG of 7! cfg=3 to 6 is typically best
The training dataset is a subset of https://huggingface.co/datasets/ppbrown/danbooru-cleaned
More details on that down below
Backstory
From an idealist standpoint, I wanted to create a finetune that was maximally compatible with everything. So I always tried to train only on SDXL base model.
However, doing this round of finetuning experiments, I was getting HORRIBLE output from base sdxl.
The colors were fine, but for some reason, the human likenesses were waayyy off.
So I abandoned my earlier principles of compatibility for now, and went for a model that does humans better: Juggernaut.
Sad story: turns out, the latest version of juggernaut actually does anime pretty well :-(
But the huggingface "Juggernaut XL" model i was playing with did not, so from where I started, it is still technically an improvement. (it is juggernaut 9)
Journal
I log some of the details of my journey, at
https://civitai.com/articles/7521/my-saga-of-fine-tuning-sdxl-models
Settings
I used OneTrainer on an RTX 4090, with the following high level settings:
input type: bf16
training type : bf16
LR: 1e-05, constant. Batchsize=1, 100 epochs
EMA step: 100, EMA on CPU, XFORMERS attention, Gradient checkpointing on.
NO text model training
I set resolution to 1024 in training settings, but did NOT enable resolution override in the "concept" definition
These settings gave me a speed of around 2IT/s and used around 16GB of VRAM
Full dataset details
You can use the scripts in
https://huggingface.co/datasets/ppbrown/danbooru-cleaned/tree/main/SUBSETS
using the input file at
https://huggingface.co/datasets/ppbrown/danbooru-cleaned/blob/main/SUBSETS/files.onegirlscene605.txt
This will download the relevant .jpg files, AND the required .txt tagging files
License
Reminder: this model is Juggernaut based, so it is subject to the juggernaut license!!