Disclaimer:

This is a character lora which is trained on NSFW material, and that is intended to produce nsfw material aswell.

warning: The header of V1 section contains Suggestive content, click "show more" at your own risk. (The majority of information can still be retrieved in the version notes)

V1:

It's finally time. I think i've finally mastered illustrious training aswell (maybe, maybe not).

This version is intended to help in the production of Mako-chan images from Mako-chan kaihatsu nikki serie.

The training has been done on a dataset of 100 images, with few repeats and a total of 8 epochs.

The dataset contained the standard sdxl resolution and so the lora works great with any of the following:

768*1344   <--> 1344*768
832*1216   <--> 1216*832
896*1152   <--> 1152*896
1024*1024

In addition to those standard resolution it has been tested on 640 x 1536 nad 1536 x 640; and it seems to work properly even in those extreme cases (the header of this section has been produced like this)

The activation keyword is:

Makochan (Mako-chan kaihatsu nikki)

it is important to note that compared to beta version there's a weaker implication of tags, so to produce a consistent output you might want to use:

short hair, bangs, hair between eyes, purple hairs, purple eyes

The dataset has been updated to reduce the problems caused by manga images, so the tags: "bubble speech, text, moaning etc etc." are no longer needed in the negative.

The Lora is now capable of a better flexibility when it comes to clothes.

It has just a little bias towards "collared shirts" but if you don't desire that kind of occasional output, you might simply put this in negatives.

To get the standard outfit you can use the tags:

School uniform, yellow cardigan, white shirt, striped necktie, blue necktie, pleated skirt

notes: using this "recipe" it might occour that the cardigan comes out as an "off-shoulder cardigan" if you don't like it, you can fix this using this tag in the negatives.

The majority of testing has been done on Hassaku Illustrious, with the following parameters:

cfg:~6
steps: 20~30
using most of the time "euler a" and "dpm++ sde"

The outputs showed here and in the gallery were upscaled and even if it wasn't needed they were clean using adetailer, to achieve an even more pleasing output.

The suggested weight for this lora is around: 0.8~1.

In most of cases 1 is perfect, but there are some nieche cases where lowering to 0.8 can fix some visual problems (as stated further).

The lora works great when combined with other loras such as concept loras and style loras; The only problem that i've been able to find till now regards the use of multiple lora with adetailer. If more loras are used, adetailer might burn the image (if the strenght of loras isn't tweaked)

This lora version has introduced a keyword to help in reproducing the hentai aesthetic, but it is still sperimental and performs as desired, only in outdoors images with the tag sunset.

This was a huge time investment (let's not talk about my gpu frying), so if you enjoy this content you can support me by following, leaving a review or by uploading your content.