Resource used: https://civarchive.com/models/22530
Weight: 1.0
Trigger: Matsumoto Rangiku
Recommended Tags: blonde hair, orange hair, blue eyes, grey eyes, long hair, cleavage, wavy hair, black kimono, silver chain, pink shawl, black wide sleeves, white sash
Recommended Negatives: alternative hairstyle, Messy hair, sleeveless
Why two versions?:
Answer: I noticed that the more Epoch the LoRA had, the more realistic the output tended to look (and this was before using photorealistic tags), so I reduced the Epoch count to 6 to create a better anime version experience and kept the 13 epoch version for realism; same dataset for both.
While these LoRA don't do a perfect job, they do it good enough. (Sadly the internet doesn't have enough accurate resources, and SD still does dumb things like white sleeves and white kimono)
Note: I use a custom checkpoint: (https://civarchive.com/models/18116 + https://civarchive.com/models/22402/fantasticmix = 25%) + https://civarchive.com/models/43977/moonmix or https://civarchive.com/models/53360?modelVersionId=57724 10%
In case the CivitAI doesn't pick it up, other resources used:
Bleach Anime Style = https://civarchive.com/models/51997
FilmGrain = https://civarchive.com/models/33208
Add Details = https://civarchive.com/models/58390
Sharpness = https://civarchive.com/models/69267?modelVersionId=76092
Noise = https://civarchive.com/models/13941
Negatives:
badv5 = https://civarchive.com/models/16807
ng_deepnegative_v1_75t = https://civarchive.com/models/4629
verybadimagenegative_v1.3 = https://civarchive.com/models/11772
By bad artist = https://civarchive.com/models/5224
EasyNegative = https://civarchive.com/models/7808
Unspeakable Horrors = https://civarchive.com/models/4499?modelVersionId=5139
Bad Prompt V2 = https://civarchive.com/models/55700/badprompt-negative-embedding
Bad-Hands-5 isn't on civitai
VAE = https://civarchive.com/models/22354?modelVersionId=80518
Description
This version is better used for Anime
FAQ
Comments (13)
so far 2/2 for awesome LoRAs, keep it up mate!
Lol, i was trying to come up with some Matsumoto images today, and wasn't pleased with the generations, so i cooked this up.
Very nice. In the future though I might suggest removing the tags for her hair color, hair style, and eye colors. Seeing as these are intrinsic to the character and should be present in all generations, they shouldn't need to be prompted. Leaving the tags out makes sure they show up without any issue since it's part of the character. Things like the clothes and hair style aren't, and are things you'd want to change. It also makes generating the character simple, meaning all you have to do is prompt the clothes and pose, instead of describing the character (which is what the LoRA is for).
Essentially, when you have those tags present in the training data, you're not giving us Rangiku, you're giving us a cosplay that just happens to look like her if you prompt it correctly.
No doubt the hair and eyes will show up even without prompting for it since it's present at large volumes in the training data, but an end user shouldn't need to prompt for it if things are tagged properly.
But yeah, good work and keep it up!
Good to know! Thanks. Will keep in mind for future loras!
@SomebodyOnceToldMeh Yeah no problemo! It's something I got tripped up with when starting myself, so I figured I'd offer up some (hopefully) helpful advice!
Knowing what to tag and what not to tag is an important part of making a good LoRA, and it's definitely something that's a skill in of itself! Best of luck in the future. Gonna have some fun with this one in the mean time!
I'm not very knowledgeable about lora, but I liked having the "orange hair" or "blonde hair" tag.
I have a scene that constantly appears with black hair, probably because I have tagged "black kimono" on it.
But thanks to the tag, you can put your color.
If the "blonde hair" tag did not exist, the AI would be changing from "blonde color" (variations) however with the tag, the character's hair color is maintained
Note: Possible spelling errors due to the translation from Spanish to English
@juanluix Your problem is literally a result of having the hair colors being tagged in the first place.
Because of how it was tagged, you have to put the hair color because as far as the AI knows because it's tagged, the hair color can be changed.
If it was trained without tagging the hair color, it would be engrained into the model and it would always assume you want blonde/orange hair without calling it up. Meaning that you prompting "black kimono", wouldn't confuse the AI, because the hair isn't something the AI was trained to change.
You don't have to use the color prompts on the tagged version, and in all likelihood it will come out with blonde hair regardless of whether you prompt for it or not (simply because all of the examples in the dataset have blonde hair), but tagging (or I guess in this case not tagging) it properly will reinforce specific character traits that make a character that character, and prevent edge cases like the one you've described.
Nice work!
nice model , hope got Saifon too 😍
this is a fantastic checkpoint you've got as well as an amazing lora. Just to clarify, is the checkpoint a merge of Citi and Fantastic with Fantastic being weighted at 0.25, and the resulting model being merged with galenda redux, with galenas weighting being 0.1?
Thank you, that is correct.
nice lora
Nice, I have increased the likes to 666
XD
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.









