LoRA of Grass Wonder (umamusume). Put LoRA file in your stable-diffusion-webui/models/lora folder and write LoRA notation in prompt to applying LoRA. <lora:your_loha_name:weight>
All of my works is free. If you'd like to support me, feel free to buy me a coffee. 👍
If you generate image of a horse girl riding horse, a horrible thing comes.
I tried many times to generate preview image. :P
Recommend options
LoRA weight 0.6~0.8
Trigger words
base
grass wonder \(umamusume\)
Racing competition clothes :
white sailor collar, blue jacket, long sleeves, armband, white dress, white skirt, black pantyhose, boots, brown footwear, mismatched footwear
(optional) naginata, holding weapon
Casual clothes :
white shirt, blue skirt, puffy short sleeves
(optional) jacket, cardigan, open clothes, long sleeves
Fantasy costume :
hat, beret, green headwear, white gloves, white dress, long sleeves, official alternate costume, white thighhighs
(optional) holding staff
sometimes image of missing ears appears. then add horse ear
Tracen school uniform :
white thighhighs, tracen school uniform, purple shirt, pleated skirt, puffy short sleeves, white skirt, puffy sleeves, summer uniform, frilled skirt, sailor collar, sailor shirt, miniskirt, frills
Use animefull based model
Settings
DPM++ Series(SDE Karras, 2M Karras, etc.)
about 20 steps, CFG scale 3.5~6.5
kl-f8-anime2.vae
CLIP skip = 1
Use highres.fix to get higher quality image
Upscaler : Latent series
Denoising strength 0.50~0.65
LoRA training info
trained on sd-scripts by kohya_ss. Thank you a lot!
based on Animefull-pruned
136 images of grass wonder
4x low quality 44 images
8x medium quality 41 images
16x high quality 51 images
=> 1 epoch = 1320 images
replaced character feature tags with grass wonder \(umamusume\)
horse girl, horse tail, brown hair, blue eyes, etc.
using aspect ratio bucketing
resolution : 768x768
rank=alpha=16
using dadaptation
batch size = 4, 10 epochs trained
clip skip = 1
This LoRA may be compatible with animefull-based models (ex. AbyssOrangeMix Series...)
All uploaded images are generated by AOM3A1
Description
first version
FAQ
Comments (24)
thanks
dadaptation learning_rate? text_encoder_lr? unet_lr?
1.0 0.5 1.0
Can I add your discord?I want to ask some question about lora training.
I have no discord. Feel free to ask questions here
what's "using aspect ratio bucketing" and "rank=alpha=16"meaning?I don't see the setting in kohya_ss.
@power70521 aspect ratio bucketing is bucketing option, argument is --enable_bucket. And this option make images auto-resizing and auto-cropping with proper aspect ratio.
Rank means Network dimension, the reason that network dimension is rank is because LoRA use low-rank dimension decomposition to approximate whole network into multiplication of two low-rank matrices.. Related argument is --network_dim=16.
alpha is a external multiplication of two matrices. Related argument is --network_alpha="16"
So, "rank=alpha=16" means both values are 16.
I hope this answer is helpful. :)
And https://civitai.com/models/12866 here are training info in training data. same as this LoRA, just different name.
@mht thanks!It's really helpful. How did you add tag for every image?which tool did you use?I use webui to preprocess images but I don't know how to add some tag to every deepbooru like "grass wonder \(umamusume\)".
@power70521 Yes, I add tag every images. But not all manually. I use below two extensions.
1. https://github.com/toriato/stable-diffusion-webui-wd14-tagger WD 1.4 Tagger.
- I think this is the best tagger from now.
2. https://github.com/toshiaki1729/stable-diffusion-webui-dataset-tag-editor Tag editor.
- remove duplicate tags(ex. tail, horse tail) and add character tag(ex. grass wonder \(umamusume\) or missing tags.
Two refining preprocess is the all of training.
@power70521 In addition, some images need to re-draw. For example, some image has unnecessary particles or backgrounds. I use Photoshop to refine images. But this process is a subjective thing and an empirical thing.
@mht thanks!!I'll try to train some lora model. and I also hope you can update satono diamond!
@mht D-Adaptation will be used after I use pip install it?or I need to do other thing?
@mht I saw a people ask about dadaptation learning_rate,text_encoder_lr and unet_lr.and you answer 1.0 0.5 1.0 and it is much different from the preset value. so I want to check again,you use these value?
@power70521 Yes, for using D-adaptation, you should install dadaptation package on venv.
Second, D-adaptation use lr as just fractional factor. D-adaptation use 'd' value in optimizer and 'd*lr' as usual learning rate. more details in https://github.com/kohya-ss/sd-scripts/issues/220.
Third things are also right. I use lr, unet lr, textencoder lr as 1.0, 1.0 0.5
@mht How to set D-adaptation value?Can I ask your value for D-adaptation?
@power70521 I use D-adaptation these args :
--optimizer_type="dadaptation" --optimizer_args "weight_decay=1.0" "decouple=True"
learning rate args
--learning_rate="1.0" --text_encoder_lr=0.5 --unet_lr=1.0
lr scheduler args
--lr_scheduler="constant_with_warmup" --lr_warmup_steps="value of 5% of total steps"
@mht thanks!Which graphic card you used?I can't use 8bit adam when I use D-adaptation?
@power70521 I use RTX3080. No you can't. 8bit adam is optimizer and also D-adaptation is a kind of optimizer. You cann't use two optimizers at same time.
@mht Have you prepared trainingData for another character?Maybe I can help you training some model.I use RTX4080 and now I'm trying to use your trainingdata to train a model like yours.I'm sorry for my bad English :(
@power70521 Thank your kindness. However, there is no prepared dataset.
Because, refining dataset has longer time than training LoRA. :(
@mht :( Really,It's hard for search lots of image.What criteria do you use to classify these image for 4x 8x or 16x?and where did you download Animefull-pruned model?
@power70521 animefull-puned is here. https://huggingface.co/a1079602570/animefull-final-pruned/tree/main
IMO,
4x is that images has 1. lower color balance, 2. monochrome but lack of details, 3. lower character's feature(ex. hair bow, eye color, etc.) 4. duplicate motion or angle.
When classification done, next step is 16x.
16x is that images has 1. adorable or gorgeous things. 2. monochrome but many details. 3. plenty of character's feature. 4. unique motion or angle.
then the other images are 8x.
@mht It's novelailatest-pruned.ckpt?
@power70521 Yes
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.