Hello again! My second ever character LoRa is here, this time featuring a character from a new Game which I can't play myself: Jianxin from Wuthering Waves. As this is a new IP, there wasn't really much to train off of and there will likely be a v2 when I'm able to create a larger dataset for training. I found every resource I could reasonably use though and made it work. Below you'll find how it was trained, recommended settings, and usage instructions.
**Note that any sample images (Aside from the thumbnail) haven't been corrected or cherry-picked to show exactly what you can expect from your generations.
TRAINING:
I manually captioned ~65 images (Can't recall the original number but they were all duped and cropped at different places to properly train clothing and individual features). There's no NSFW images in the dataset, but you can still create works of that nature with proper prompting and LoRa weight choice; I trained this LORA using CivitiAi's provided training service in 768 resolution and more repeats than normal so the clothes were accurate.
Recommended Generation Settings:
Trigger Word: Jianxin \(Wuthering Waves\)
Recommended Prompt(s): Jianxin \(Wuthering Waves\), 1girl, solo, silver eyes, multicolored hair, black hair, white hair, ahoge, yin-yang earrings, tassel earrings, low twintails, cube hairties, china dress, black and white dress, Tacet mark, frequency tattoo, armpit cutout, white straps, asymmetrical sleeves, white hand glove, black elbow glove, green details, white belt, black shorts, green lining, asymmetrical legwear, single thighhigh, high heels
**Remove keywords from the prompt as you see fit
Settings:
Dimensions: 512 x 768, 768x768, and 768x512
Sampler: Euler a
LORA weight: 0.5 - 1.0 (I recommend 0.7 or 0.8, lower it for more variety of backgrounds)
Recommended Checkpoints: To my knowledge it works with any Pony checkpoint but here's what I like using:
AutismMix - https://civitai.com/models/288584?modelVersionId=324619
MatrixHentai - https://civitai.com/models/356456?modelVersionId=445225
For Txt2Img, 512x768 and matching aspect ratios is recommended. You can do landscape generation with the inverse ratio of 768x512 or square ratios if another LORA should require it, but some features may not be generated as consistently without inpainting. This LORA is very consistent with facial features with weights at 0.7, higher weight will help clothing accuracy but may also skew jewelry or hair. It's also pretty consistent with bottomless and completely nude generations, just add the common Booru tags relating to those scenarios as needed.
Please leave feedback and share your generations! I'd like to know how I can improve this process! If you have other character recommendations, I'm open to suggestions if any pique my interest. As always thanks for supporting, Buzz tips are appreciated (I'm -1500 after training this one lol) so I can train other character and anatomy LORAs for my own works, and check out my Twitter/𝕏 or Patreon if you feel like supporting :)