Yarai Miu from the visual novel Dracu-Riot!
Main tag is yarai miu and and her braided hairstyle with the ribbon can be set with ribbon braid.
There are 4 outfits trained:
public safety bureauThe military style work uniformfancy outfither purple casual outfit. Does not mess well withribbon braid, it can hilariously start adding ribbons on the "braids" on her shirt.aqua eden school uniformthe school uniformsleepwearher green nightgrown. Does not work well due to lack of training data, but you can get something similar withgreen sleepwear, short sleeves, shorts
This has been trained on fanart and game event CGs. I might try adding the character CGs to improve the outfits in the future.
Description
FAQ
Comments (10)
You re the best.
Miu best girl
Whatever you did with your lora seems to have insane prompt adherence, great work! Would you mind sharing the training settings you used? I'm currently experimenting to see what gets the best balance between prompt adherence and style fidelity
As for adding CGs for the different outfits, I found that just having the full sprites cropped to the full body helped a lot, as well as rigorous tagging and prompting. With my Elina lora, I only had 3 images (two sprites, 1 cg) for some outfits like the pajamas and bikini, and they both seemed to pop up fine with prompting.
Compared to your lora, I try to learn outfit tags. So I take all the related clothes tags for an outfit, remove them and replace it with a single custom tag. That is how I want it to work, so that it recognizes specific trigger words instead of slightly modifying common tags. I want the prompting to be simple and I also want to move towards a future where you could combine the Lora's into a larger model.
But you need to be careful that the tokens it prompt translate to doesn't overlap with other tags. It probably also require more training images to work well, though there was a lot of fanart for Miu (compared to other visual novels). I use a regularization data set of about 190 random images in an attempt to prevent it from changing ordinary tags and focusing on my trigger tags. There is a different training approach out there where you learn an textural inversion embedding together with your lora which I want to try.
I tried the Mitsukasa Ayase lora which is here on CivitAI and that had very clearly over fitted on the character sprite pose (as well as having used transparent PNGs during the training with garbage in its transparent areas), so I want to try to be a bit careful. (But I didn't include them this time mainly because I had them with the body and faces sprites separated in the CG dump and I hoped it would be good enough without them.)
Ahh I see, can you upload your dataset sometime so I can sift through it and test to see if some tagging differences and regularization would allow for higher flexibility in some other loras I'm working on?
Also with the sprites, in my experience it shouldn't matter as long as you've got a large enough dataset, with 42 images (11 of them being sprites, very useful for details you don't see in the CG), I didn't notice any overfitting as long as they were properly tagged (in my case, black background, simple background, full body, from side with some).
Also yeah I tried some other Yuzusoft loras that I want to get around to rebaking for similar reasons. The Inaba Meguru one was trained on Abyss Orange Mix and had very little tagging, while the Arihara Nanami one seems to have had inconsistent tagging or maybe even overtagging which resulted in prompting for outfits like the school uniform to come out as a different school's half the time.
@5e5 I will not be making the dataset publicly available since it is only containing copyrighted fanart/game art. I don't mind sharing in private however. But it is basically autotagged, then I manually tag the characters and outfits, and then filter out the automatic tags based on the manual tags. So the tagging is quite lazy.
I'm not particularly a YuzuSoft fan, so I only plan on working on Miu and then Ayase from Riddle Joker (I just added my first version today, not quite satisfied yet though).
You are my hero!!!!!!Hope you can make more Yuzusoft characters!!
Sorry, from Yuzusoft I only plan on working on Miu and then Ayase (Riddle Joker)
Thank you for the great model!
Can I also ask for mera azusa?
Sorry, I don't take requests, I only work on characters I have an interest in myself. But on the other hand I don't beg for donations or Patreon subscriptions : )
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.