As usual, the model is trained with clip skip 1 on top of ACertainty and locon is extracted from the trained model
The locon should work on most models as I show in the xy plot
All the intermediate checkpoints can be found on https://huggingface.co/alea31415/shokei-shojo
It includes the following 7 characters from the series
Ashuna
Flare
ManonLibelle
Menou
Momo
Pandemonium
TokitoAkari
For Menou you have both
Menou
Menou child
For Pandemonium you have
Pandemonium (the classic one)
Maya Pandemonium (hair down version, the one after the end of the 5th tome of the novel)
Centipede Monster (monster version, the one at the 12th episode of the anime)
Maya Pandemonium is only trained on 6 images from the light novel: 2 colorful and 4 monochrome, but it should still be usable if you adjust things properly
The preview of Centipede Monster is only included for the FullCkpt version, because I guess it may make some uncomfortable (and anyway I do not think anyone would ever use it)
I split the styles into 3+1 during training
aniscreen (3928 images)
fanart (95 images)
colorful light novel / light novel (153 images)
As usual I use a lot of regularization images during training, in particular I also include most of my dataset for the in container loha (https://civitai.com/models/66021) here so I guess these concepts are more or less learned as well
The example images are either generated by the LoCon on top of VBP2.3
Or using a certain merge of the trained model
None of the images are generated in one shot
They generally include a lot of random draws, img2img, inpainting, and sometimes use of controlnet and regional prompter (for the case of multiple characters)
--
Just to complain, I really have a hard time figuring out what kind of images violate civitai TOS
I wanted to upload an image with Pandemonium on top of ManonLibelle who is lying in the blood (this one), but apparently this is detected as violation of TOS and gets removed automatically
In the mean time, we can easily find a lot of images that are on civitai that are more violent/horrible/bloody or whatsoever
Otherwise, although probably few people would ever be interested in a model for this
Shokei shojo is basically my favorite light novel, so I kind of feel that I must do it
Btw I saw an image of Menou from this discussion on outpainting of controlnet https://github.com/Mikubill/sd-webui-controlnet/discussions/1464 so I told myself it was time to get this model done