Hey everyone, I've been looking everywhere for a LORA for V7-V9 Weiss and never found one and sort of had a "fine I'll do it myself" moment and here we are. This is my first character LORA and it became apparent that the likely reason this LORA doesn't already exist is because there isn't an abundance of works featuring this version of Weiss. 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 haven't been corrected or cherry-picked to show exactly what you can expect from your generations.
TRAINING:
I manually captioned 140 images (76 unique pictures with most duplicated then cropped at different places to properly train clothing and individual features). Most of the SFW resources utilized concept are or anime screenshots and ~20 of the 76 were NSFW artwork references; I trained this LORA using CivitiAi's provided training service in 768 resolution. I couldn't modify the steps used and didn't want to over-train by adding too many epochs since I and I'm sure many of you have various intentions with this LORA, and that would skew it's use cases. As a result, I opted for a lower learn rate so features like her scar and the jewelry were properly retained as the LORA was trained without over-training. Bucketing was used so it supports multiple resolutions.
Recommended Generation Settings:
Trigger Word: WeissVol7
Recommended Prompt(s): WeissVol7, blue eyes, scar across eye, white hair, braid, high ponytail, large braid, braided ponytail, long ponytail, sidelocks, hair ornament, crown ornament, earrings, jewelry, red earrings, blue-trimmed dress, blue collar, puffy sleeves, cleavage, snowflake insignia, black gloves, white belt, pouches, combat dress, high heels, boots, wedge heels,
Settings:
Dimensions: 512 x 768
Sampler: DPM++ 2M SDE Karras or Euler a
LORA weight: 0.7 - 1.0
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 upper body features with weights greater than 0.7, if you have issues with getting the dress or skirt to show up, use parentheses around them to fix it. 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 so I can train other anatomy LORAs for my own works, and check out my Twitter/𝕏 or Patreon if you feel like supporting :)