V2:
Wow this one was old. I retrained it to current standards and Isolated her maid uniform. I also made a LOCON version with the same exact dataset and similar config values so download both and compare. It is nice how they are so similar and different at the same time. Anyway...
Works fine with: <lora:Maria_VMV2:0.7>, Maria_VM,
Her maid outfit can be triggered with:
VIctorian_Maid_Uniform
Also If you wonder about her maid bonnet(? or whatever it is called). I folded it into her character trigger. I just don't have any image of the top of her head without it.
As previously the dataset is mostly screen caps as there's not much original art. So no dataset for you, instead I will add the Tags summary. There's some interesting tags in there like her unique garter_belt. So check it and experiment.
V1:
Fell down a rabbit hole of debauchery and made a LORA from one of my favorite H anime. Seems to work fine from .4 to .7 and gets a bit cartoonish at higher levels.
Got good results with: <lora:Maria_VM:.5>, Maria_VM, big_breasts
Description
FAQ
Comments (3)
Was there any horizontally skewed images in your training set? On some generations I am seeing what looks like "stretching to fit" images on 512x768 and 512x704 images, where the generated image looks too narrow. It doesn't always happen, but sometimes does: https://imgur.com/a/fJRycqQ
I am using corneos7thHeavenMix_v2, and a few other modifier LoRAs (but I normally use those on your other LoRAs as well)
It might also just be the normal art style, and not a real problem? Other than that, models works well.
Edit: Now that I've used this LoRA a few more times, it seems higher weighting of other LoRAs can correct/alleviate the skew look.
I think I might know what caused that. Found a convenient gif pack with facial expressions. So I split them. Problem is they were 600x400 so I cropped them and filled them up with white bars at the top and bottom, no upscale though. Also I do all my training in 512x512, that could also be part of the issue.
Edit: BTW Your images look a bit over saturated, you might be using too high a weight or it could be a VAE or model issue. This days I train everything in AnythingV4.5 It is not the best but it seems a good enough "Standard".
@knxo I think the oversaturation is coming from the One Piece wano style LoRA, or could be the vae like you said. I don't really mind it. It could also be coming from one of the other 3/4 LoRAs I was using haha.
I've used every single one of your LoRAs, and they all work very well. I think your training method is a great standard.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.