LoHa of Mejiro Ardan (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. 👍
Recommended options
LoRA weight around 0.7
Trigger words
base
mejiro ardan \(umamusume\)
Racing competition costume :
ear ornament, black gloves, white gloves, asymmetrical gloves, mismatched gloves, detached sleeves, puffy sleeves, short sleeves, long sleeves, asymmetrical sleeves, uneven sleeves, ribbon, bare shoulders, off-shoulder dress, black dress, frilled dress, center frills, knee boots, asymmetrical footwear, mismatched footwear
or simply. mismatched gloves, mismatched footwear
Asymmetrical thing is very very hard for SD model.
SSR Speed :
official alternate costume, ear ornament, choker, detached sleeves, wrist cuffs, off-shoulder dress, white dress, strapless dress, bare shoulders, collarbone, blue thighhighs
(optional) hand fan, holding fan, folding fan
Due to lack of data, It reproduce rarely. sry
Casual :
jewelry, necklace, long sleeves, white shirt, belt, necklace, blue skirt, long skirt, floral print, socks
Tracen summer uniform
tracen school uniform, summer uniform, serafuku, sailor collar, bowtie, puffy short sleeves, sailor shirt, purple shirt, white skirt, pleated skirt, miniskirt, white thighhighs, brown footwear, shoes,
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
Upscaler
Latent, Latent (nearest-exact), Latent (bicubic antialiased) or other 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
136 images of mejiro ardan
4x low quality 46 images
8x medium quality 45 images
16x high quality 45 images
=> 1 epoch = 1264 images
replaced character feature tags with mejiro ardan \(umamusume\)
1girl, horse tail, horse ears, animal ears, tail through clothes, blue eyes, horse girl, very long hair, purple eyes, tail, long hair, solo, bangs, blue hair, grey hair, crown braid, simple background, light blue hair, braid, ear ribbon, hair ornament, off shoulder
I think these tags are
the essentials of mejiro ardan
or unnecessary for training because of duplication.
resolution : 768x768 with aspect ratio bucketing.
other training settings is included in training_data.zip or metadata in LoRA file.
All uploaded images are generated by AOM3A1
And, special thanks to my best friend. DH.
FAQ
Comments (13)
Scarlet is in progress :)
Please, I want to draw bare legs in racing competition costume. Even if I write "bare legs" it's shiny.
If you copy and paste prompt of preview image, it is because of glassed legs tag
赤い髪のキャラクターが出力します。使えませんね
A character with red hair will output
Did you check LoRA notation in your prompts? If you copy prompt of preview images, it may differ in LoRA name.
@mht Isn't it mejiro ardan \(umamusume\) ?
@mht I was able to print thank you
Thanks for your great work, your model works well.
I'm a beginner of LORA training, may I ask you two questions?
1. Did you add some regularization img when trining? If so, what is the ratio of dataset to regularization?
2. Your images are divided into three quality types, how did you do that? Any specific tools?
And is it important to train with image quality classifications?
I would be appreciated if you could reply, thanks again.
1. I prefer no regularization images. Because regularization needs for preserving original model(base model). With LoRA, just to turning off can preserve base model. I think regularization is only needs for full fine tuning. LoRA is not full fine tuning method.
2. Divining images by quality has a benefit for training. By controlling repeat number, LoRA model learns more about higher quality than lower quality images. And there is no specific tools for separating images. I scan every images and score each image(in mind).
Here some standard for scoring image I use.
1. Image which has details of character get high score.
2. Whether Image is monochrome or not, if image which has details of character get high score.
3. If image which has good balance of colors, It will get high score.
4. After intrinsic scoring done, I separate image which has lower score to low quality folder.
5. and then, I move image has higher score to high quality folder
6. Left images go to medium quality folder.
@mht
Thank you so much for such a detailed imformation !
I believe this will help me train my model
so glad to receive your reply
Please add dance costumes
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.

