LoRA of Daiichi Ruby (umamusume). Put LoHa file in your stable-diffusion-webui/models/lora folder and write LoRA notation in prompt to applying LoHa. <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. 👍
Recommend options
LoRA weight around 0.8
Trigger words
base
daiichi ruby \(umamusume\)
Racing competition costume :
red ascot, puffy long sleeves, collared dress, frilled dress, purple dress, white pantyhose, mary janes, black footwear
(optional) playing instrument, holding instrument, violin
Casual clothes :
casual, long sleeves, blue shirt, collared shirt, shirt stuck in skirt, blue skirt, long skirt, flat skirt, white pantyhose, blue footwear, loafer
Due to lack of dataset. It is occasionally appearing.
Tracen school uniform :
tracen school uniform, purple shirt, pleated skirt, puffy short sleeves, white skirt, puffy sleeves, summer uniform, frilled skirt, sailor collar, sailor shirt, miniskirt, frills, white thighhighs
Use a model derived from or mixed with animefull model.
Settings
DPM++ Series(SDE Karras, 2M Karras, etc.)
about 20 steps, CFG scale 3.5~6.5
I recommend kl-f8-anime2.vae
CLIP skip = 1
Use Hires.fix to get higher quality image
Up-scaler : Latent series
Denoising strength 0.50~0.65
LoRA training info
trained on sd-scripts by kohya_ss.
Thank you all a lot!
based on Animefull-pruned model
44 images of daiichi ruby
4x low quality 10 images
8x medium quality 15 images
16x high quality 20 images
=> 1 epoch = 480 images
replaced character feature tags with daiichi ruby \(umamusume\)
horse girl, horse tail, drill hair, pink eyes, etc.
resolution : 768x768 with aspect ratio bucketing.
other training settings is included in training_data.zip
This LoRA may be compatible with animefull-based models (ex. AbyssOrangeMix Series...)
All uploaded images are generated by AOM3A1
And, special thanks to my best friend. DH.
FAQ
Comments (17)
What's your impression of the LoHa method?
I use same rank after calculating Hadamard product with LoCon's rank.
(LoCon rank 16, 37.2MB = LoHa rank 4, 12.07MB)
Just using same as rank of LoCon make model more complex and overfitting.
Considering this difference btw LoCon and LoHa, LoHa has at least equal quality and less file size.
So, I think It is good at both quality and file size, but more complicate for makers :)
In addition, LoRA(LoCon) of Ruby I made has been totally broken. I think It comes from dataset but I don't know how to fix and re-train.
What a great work!
Can I mention your model in my repo?
Yes, you can👍
I can't use your LyCORIS model normally. Even if I install the lycoris extension, I don't know whether there is a problem with my installation or other problems. I can't use your current Daiichi Ruby and the latest mihono lycoris model normally. I was able to work with locon's mcqueen model before, which is weird
Did you update locon extension? Locon extension of my version is 8aa2b095 (Mon Mar 13 05:09:30 2023).
@mht Yes, my locon extension is the latest 8aa2b095 (Mon Mar 13 05:09:30 2023), maybe I didn't get some details right
@toki233 Is there any error or warning message on WebUI console? Something like 'loading LoRA fails,[LORA_NAME].
@mht Unfortunately, there was no error. After that, I tried to delete the lycpros and locon plug-ins installed on the network, and used the directly compressed package to download and then put into the extensions to decompress. The result was that the lycoris model of mihono and ruby could not be used. locon's myqueen works normally
Or are the lycoris and loha models actually two models because I didn't install the loha model causing an error?
@toki233 2889897157914 said that just putting in loha in SD\models\lora folder solves the problem. Is it works?
And LyCORIS is a package for training. This package is not necessary for generating image.
Yes, I also tried putting lycoris into models\lora, but it didn't change anything,Are you saying that lycoris is only required for training models, and that using lycoris models requires loha extensions?
@toki233 A model of creating by LyCORIS and type of LoHa needs locon extension when generating images.
In short,
LyCORIS -> Training pakage
A model by LyCORIS -> Requiring LoCon extension for creating
@mht Ok, I tried again with A1111-Sd-wewe-locon
Uninstalled and installed again because the locon extension also contains loha? I also tried to download and unzip this file into modle\lora. The result is still the same. The loha of ruby cannot be used and the locon of mcqueen can be used. I am a little confused
@mht I have already had the idea of giving up. Thank you for your answer and I'm sorry for taking up a lot of your time. Let's end this matter here for the time being. I like your character model very much, wish you better and better
@toki233 In now, there exists three names of LoRA.
LoRA, the origianl.
LoCon, LoRA for Convolutional network. Extended version of LoRA.
LoHa, LoRA with Hadamard product. This has LoCon architecture, only difference is using Hadamard product for layers in model.
In theoretically, these are ALL LoRA. All of these use Low-Rank Adaptation method from microsoft.
So, locon extension is for LoHA and LoCon.
And lastly, I'm sorry 😢
The additional network does not support LyCORIS now (2023.3.15). So, I can just put the LyCORIS in the folder:D:\SD\models\Lora. It works, but using the LyCORIS can not get a good effect.
This is one big PROBLEM of LyCORIS.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.