This is a LORA based on the character Shiomiya Shiori from The World God Only Knows. This Lora does good results around 0.6. This model was trained on WarriorMama777's orange mix from huggingface. Made with the Kohya tool for AI Training. Any suggestions to improve or recommended things for the data set would be appreciated.
Description
Do to updates of the training tools I use I had to change my usual training method. If this shows to be a problem, I will try to change it again though it seems to be working as usual in my test.
FAQ
Comments (8)
夕宫栞yyds
I need instructions, I just started making Lora an anime character, I have 25 cropped screenshots of the Staz Charlie Blood character. But I failed after trying to train it. I want to know your settings. What must be set so that it can recognize the character's face?
Before anything crop your images on Birme to a 512 x 512 size then tag your images in the danbooru format (it can do it automatically in automatic1111 but definitely fix it up by hand and put the term your teaching in the front of every prompt .txt file). After that the rest is in the kohya training tools.
I've personally moved to the abyss orange mix 3 checkpoint but train for whatever you use for best effect.
First go to the lora training tab and place the location of your ckpt in the 'Pretrained model name or path' box. Below are all the training settings I use. It's not perfect but it gets working results from my experience.
---Training Parameters---
Train Batch Size:
2
Epoch:
1
Save every N epochs:
1
Caption Extension:
.txt
Mixed Precision:
bf16
Save Precision:
bf16
Number of CPU threads per core:
2
Seed:
1234
Cache Latent: on
Learning rate:
0.0001
LR Scheduler:
constant
LR warmup (% of steps):
0
Optimizer:
AdamW
Text Encoder learning rate:
5e-5
Unet learning rate:
0.0001
Network Rank (dimension):
128
Network Alpha:
64
Max resolution:
512,512
Stop text encoder training:
0
-Advanced Configuration-
Gradient accumulate steps:
1
Prior loss weight:
1
Keep n tokens:
0
Clip skip:
2
Max Token Length:
75
Memory efficient attention: on
Use xformers: on
Don't upscale bucket resolution: on
Bucket resolution steps:
64
Dropout caption every n epochs:
0
Rate of caption dropout:
0
---Tools---
-Dreambooth/LoRA Folder Preparation-
Instance prompt:
(whatever it is)
Class prompt:
(general version of what it is ex: character, landscape clothes)
Training Images:
(your images)
Repeats:
200
Repeats:
(1)
(repeats there is one for training and 1 for regularisation images)
Destination training directory:
Wherever it is
@rulles thankyou for sharing you knowledge, by the way, how many images can you use for settings like the one above?
@satoshipanutanku861 From my experience my Oosuki Mamako Lora is a bit overtrained from the look of it. That had 132 images, a good minimum is 15 images with that in mind your ideal image range is between 15-100.
@rulles I want to ask again, how can AI know the character's school uniform? Do you write the word "school uniform" into the txt tag and caption file?
@satoshipanutanku861 yes tag it for school uniform and put that tag near the front, right after the character name ex: name, school uniform, and so on. Also make sure it has several examples of the character wearing it 1 or 2 is not enough varies on total data set size.
@rulles It’s a rare thing to find someone as generous and kind as you. Thank you.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.