I continue evading the cleaning of that new dataset. Instead i dusted this one, cleaned it again, added some extra costume images and re tagged it. Thus this shares the dataset with my Chichi V2 Lora
Works fine with: <lyco:chichi_dbzV3:0.7> chichi_dbz, <outfit>
I added triggers for the following outfits:
Purple_Dress_Orange_Neckerchief: Variation of her normal attire in purple and red.
Chichi_Bikini_Armor: Chichi cosplaying as her self in her Bikini armor.
Chichi_Blue_China_Dress: Her DB tournament arc dress look alike. I don't remember her using it during her MILF days but i found images and i liked it. Her MILF version of the attire is pants less of course.
Yellow_China_Dress_Purple_Neckerchief: Variation of her normal attire in yellow and purple. This probably need to add Neckerchief as extra or in negative as most of the images didn't have it and it is iffy. Sadly couldn't make this one consistent as it seems to fight against the top of her dress and turns it into a ribbon.
As Always I'll add the dataset so people can check the tag summary and experiment.
Description
FAQ
Comments (4)
great butI have a question, I saw that you uploaded a lora version and this lyco and apart from the difference in format, what other difference is there between the two that you uploaded?
Basically the locon also train some other parts of the neural network. thus the con part which means the convolution layer, so think lora+ but very little plus. Mostly it adds slight more detail to stuff and catches a bit more subtle differences. Thus it gives a slight better concept separation (like when using blue dress sometimes makes everything blue.). Otherwise it is a bit more cross compatible between models as it exports "more". TLDR very slight image improvement, some better concept separation and a slight better tolerance to different models. For example if you used it on a photo realistic model you might have slightly better luck using the locon version. I honestly just train them because i like to see the subtle changes against the lora version and that they only take the extra baking time as i can reuse the dataset.
Could you do the teenage version that participated in the martial arts tournament?
I don't really like to repeat characters unless it is on an specific doujinshi author style(That i really like) or the current one is hideous/loli. You could try https://civitai.com/models/25868/chi-chi-dragon-ball