main point of this is that both characters are sitting up
trained on the following phrases:
upright straddle, pov, close-up: moves the face closer
upright straddle, pov, hug: girl hugs with both arms
upright straddle, pov, one-armed hug: hugging with one arm, helps to add,
leaning backandarm supportupright straddle, pov, torso grab: grabbing hips and or waist
upright straddle, pov, breast grab: grabbing the boobers
upright straddle, pov, pov-hug: pov arms behind the back
upright straddle, pov, ass grab: grabbing the butt
upright straddle, pov, face grab: grabbing the face
upright straddle, pov, holding hands: holding hands... works, but hands are kinda jank
upright straddle, pov, thigh grab: grabbing upper thigh
hug with close-up works best, one-armed hug is kinda gacha, but works great when it does show up
Description
first version, trained to 10 epochs, I'll probably upload a twenty epoch version soon
FAQ
Comments (9)
note: best at tall images, pretty bad at wide images
also! from above helps a lot too.
two versions, v1 is 10 epochs, and v2 is 20 epochs,
not sure which I prefer; let me know your results!
for v3, I'll see if I can reduce the bar censors
For me, v2 is too overtrained (?) I prefer v1 personally.
Hey, does this Lora work with realistic models? The examples here are all anime style.
@fdjw88946 I just tried it out on chillout mix, and it seems to work.
It looks like from your examples that your dataset had a lot of censored images. Can you prompt for "uncensored" and expect to get it most of the time?
not reliably in this version, but you can use
https://civitai.com/models/11830/steampunkengibarcensor-bar-censorship
at 0.5 strength, and put 'bar censor' in the negative, and it gets rid of them.
@emberbyte Oh nice! I knew there were other tools like on GitHub, but I didn't realize someone had made a decensorer for SD. Thanks. 😄✌️
This works very well and looks beautiful, but it significantly reduces the image quality.