PonyXL update (16/1/25)
I've decided to return to my roots and update the first style I uploaded. To only find out I suck at training styles in Pony.
The LORA I trained can replicate the style fine(ish) on original standalone prompts but once character loras get added to the equation, the style disappears.
Because of that I had to work on this for ages before getting some degree of resemblance.
Also this lora is a lycoris-locon.
The style resemble more of the artist's doujin cover illustrations as that's what I mainly used.
strength: 0.9-1
add empty eyes for a 'mind control/mind broken' expression
I recommend using 80% of the chara LORA's usual strength if you are going to be using characters along with the style lora.
Adetailer may help.
SD 1.5
update:
added background in training data, so backgrounds are more likely to show up instead of blurry white/dark bg.
while testing, it's has a tendency to gen a pov hand so put 'pov' in neg prompt if unwanted.
tested on grapefruit 4.1 - ideal weight for that model is 0.75 (needs more testing, will edit soon), other models may vary
also these images are pretty cherry picked, still working on the right set of prompts for good hands (can only get 1 in 10 gen with good hands with example image prompts so far)
\\\\
First time posting a lora, made some before but I wasn't happy with them.
Artstyle of the hentai artist Ma-kurou (or fyama)
artist pixiv: https://www.pixiv.net/users/1293240
weight 0.7-1.0 works fine, no trigger words needed
trained on NAI, tested on Anylora
Note: white wips of steam or breath was present in a lot of the training images,
so put 'steam' or 'breath' neg prompt if unwanted. Same with sweat and wet. There was also a lot of 'simple background' (black, white, etc) in the training data so look out for that.
I might make a v2 with backgrounds edited in but not now, I just want to move on to other things.
Description
added background in training data, so backgrounds are more likely to show up instead of blurry white/dark bg.
while testing, it's has a tendency to gen a pov hand so put 'pov' in neg prompt if unwanted.
tested on grapefruit 4.1 - ideal weight for that model is 0.9 imo, other models may vary
also these images are pretty cherry picked, still working on the right set of prompts for good hands (can only get 1 in 10 gen with good hands with example image prompts so far)
FAQ
Comments (2)
Man, I have to love you. It's strange that you were the only one who chose to create the artwork. Though I believed him to be well-known, I was even considering creating an adaptation of his artistic style. So, how did you compile the dataset (pictures) and, if you recall, how much of them did you use? I was under the impression that he drew rarely than his usual doujinshi which you have to crop a lot.
I've lost the v2 dataset but had uploaded the v1 dataset at the time. They're pretty much the same if you want to have a look, the only difference being most simple background images had their backgrounds replaced with simple scenery.
I had to dig through social media, pixiv, danbooru and hentai sites for collecting. When that wasn't enough, I included a small number of monochrome images. In hindsight I could done it in an easier way with ai upscaling and cropping.
A tool I used a lot was the retouch function when editing images in 'Photos', the windows default image viewer. Helped with titles, logos and watermarks if it's not it a vital spot and turns what couldn't be used as dataset into something it could be.