My data set for this was very small, so lots of repeats. You need to use it at lower strengths or it burns the image alot.
But. At lower strengths it does a pretty good job. Really low. Like 0.4 low for good results.
Trained on:
standing sex: standing sex
legs together: legs together
legs apart: legs slightly parted
imminent: slight penetration with whole penis visible
penetration: full penetration, penis still visible
ballsdeep: deep penetration, penis not visible
from above and foreshortening help, also works well with public and exhibitionism.
I might retrain. The first version works, but it's kinda hard to control for some of the more specific poses.
For example. legs together, and legs apart, you need to put attention on the together or the apart, but not legs. ie legs (together) or legs (apart)
For imminent, penetration, and ballsdeep? They work... ish, if you generate at 768x768, but for some reason don't seem to work at 512x768, and 512x768 has much better results overall? So not sure what to do about that.
Description
FAQ
Comments (13)
NOTE: I've got a few versions posted, let me know your results! The higher epoch versions give more control, but are more burnt. I probably need to retrain. The main issue is that the data set for these poses was very small.
My sample images are all the same generation data, except with the different loras, so you can see what it's like.
Except for image 1 in v1, for that one, I lowered the strength.
TIP! Works very well with my bulge lora. I just uploaded some pictures of that for v1.
It seems it doesn't work good with realistic models.
I just tried it out on chillout mix and seemed to do ok, the generation I did was
```
standing sex, legs together, imminent, pov, wall, against wall, alley, public, (from above, foreshortening:1.1), standing, outdoors, stealth sex, happy, smile, blush, grey eyes, white hair, (against wall:1.2), clothed sex, skirt, t-shirt, long hair, hug, legs together, [medium breasts, mature female:0.2] <lora:concept-standing_sex-pov-v1:0.4>
Negative prompt: (worst quality, low quality:1.4),
Steps: 33, Sampler: Euler a, CFG scale: 7, Seed: 3162251548, Size: 512x768, Model hash: 7234b76e42, Model: realistic_chilloutmix_Ni, Denoising strength: 0.7, Clip skip: 2, ENSD: 31337, Hires upscale: 1.2, Hires upscaler: Latent
```
at least with that prompt, the results are consistently good on chill out mix, regardless of seed. I didn't do much testing for realistic though, this is trained on nai, and I don't usually do realistic generation.
Overall a great lora. I had trouble with having the leg lift work but I did have to lower weight to around 0.4 for it not to obliterate my model style so it might be related. Also it's kind of hard not to have the dude grabbing the lady's wrists.
But all in all, very nice!
What's the difference between versions?
number of epochs, higher epochs means it gets the stance better, but burns the style more. v1 being 20 epochs of training. Generally, v1 at 0.7 strength should give good results.
I do probably need to retrain this though, to stop the style burn.
低了没效果，高了影响人物，很难受
....
This one is really amazing. Do you think it would be possible to replace to "rubbing/fingering/making squirt" pov ? Like just a pov hand coming from the camera to her crotch.
Are we supposed to use all the trigger words together or seperate? I used all together and got a standing doggy position when I wanted a front pov 1 leg lift position
过拟合了，需要降低权重来用
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.