Halide, a rank-96 DoRA for Krea 2, trained on 19,500 images.
The dataset includes anime/hentai, cinematic movie stills, photos, 3D/SFM, and AI-gen from older models. It is captioned with a mix of tags and full sentences. Dataset gathering, tooling, and training took months of work, much of it built before Krea 2 even released. Trained at home locally with onetrainer.
It's not pure NSFW, but since that was Krea 2's weak point it got the most focus. Concepts and positions it was trained on:
Standing doggystyle
Doggystyle
Cowgirl
Reverse Cowgirl
Missionary
cumshot/creampie
Lotus position
Reverse Lotus position
DP/Double penetration
Bondage (lots of different bondage kinks)
Pronebone
Blowjob/fellatio
Sidefuck
Cunnilingus
Handjob/masturbation
Some sex toys like dildos, magic wand/hitachi vibrators etc
Fingering
Fisting
Standing carry
Prompt with tags or sentences. For photos say photograph, raw photo, video still, (not photorealistic); for film-style try cinematic movie still. see the gallery.
For concepts and words normally associated with anime/booru the model can sometimes drift towards realistic-anime / 2.5D / 3D-looking images. It can help to use a two stage ksampler: first stage with the raw model and a cfg above 1 and a negative prompt. Some of my gallery images contain that 2x ksampler workflow I used, with toggle on/off.
Description
FAQ
Comments (2)
Hey thanks for this work :) Looks promising but under ComfyUI I have a whole bunch of lora keys that don't load. Anyone else have this issue ?
Thanks for spotting that! It was a format mismatch between what I picked in onetrainer and what comfyui prefers. The directional LoRA weights were loading fine, but ComfyUI was skipping the DoRA magnitude vectors.
I've re-uploaded the file converted to native ComfyUI format, so it loads all tensors with zero console warnings now. Visually it's a subtle tweak (mostly slight micro-contrast/seed shifts), but everything is 100% properly mapped now. Thanks again for pointing it out!