The dataset is growing: this new version includes 104 images, trained directly on the captions. I'm expanding the LoRA to integrate the main concepts of Fan No Hitori.
So far, I've included:
image 'safe'
tentacles
shirikodama
cursed painting
some example of boxification.
All the images in the dataset have their own captions, which you can use as a reference for testing.
The basic structure of all descriptions starts with:
"Extremely explicit NSFW digital illustration color in the clean colored digital style of fan no hitori of".
Important note: since the dataset is almost entirely monochrome, you'll need to emphasize color-related tags/prompts to avoid getting black-and-white outputs.
I highly recommend using SeeSeeLP's base model, the results are fantastic compared to Krea 2.0's base models. I also used their workflow, and it works perfectly.