Update V2:

reworked the dataset and retagged with qwen 3.5 vl. dataset size now around 800.

trained rank32 rank64 and rank128 versions for 8000 steps each. rank32 was noticeably worse, rank128 didnt improve anything

new version seems to have generally better prompt adherence.

V1:

i've experimented with combining my smaller datasets into bigger more diverse concepts with older checkpoints a lot before. this one is the first one that come out kinda useable.

this lora is able to produce rope bondage in various poses with various details

the model has strong preferences for certain poses and details: if you are looking for something specific: experiment with your prompt and if in doubt add more redundant exhaustive descriptions. I've so far been able produce everything i know to be in the dataset (and a little more).

(this could probably benefit from better consistent captions. i've run the whole thing through qwen 3vl abliterated and its mostly fine but pose language can be a bit off. if you know a better captioning model that doesn't complain to much about me being too naughty: let me know)