New Illustrious version!
I've made an article for people who are wondering how I keep my backgrounds in the videos consistent.
https://civarchive.com/articles/23205
The videos took a long time to get right on this one! Glad with how they ended up but I might limit myself to one per LoRA/update as it really eats at my time and slows me right down which isn't ideal.
With this update the LoRA should be a lot more friendly with other LoRAs and be very customizable, so please feel free to play around with prompts and see what you get! 😁
Core Prompts
for missionary
MLD, 1girl, 1boy, pov, lying down on ground, on back, straddling,
for mating press
MPLD, 1girl, 1boy, pov, lying down on ground, on back, legs up,
for doggystyle
DSLD, 1girl, 1boy, pov, lying down on ground, from behind, butt,
customization prompts
pinning, pov arms,
looking at viewer,
dutch angle,
from above,
reaching,
reaching back,
sex,
sex from behind,
imminent sex,
(cum in pussy, overflow, cum bubbling:1.1),
nipples,
breasts apart,
breasts together,
breast grab,
butt grab,
leg grab,
Description
With this update I've tried to overhaul the whole LoRA with new training pictures as I thought the effect was quite limited.
This is a pov sex position Lora with 4 main positions and customising prompts.
Core Prompts
Pov Missionary,
Pov Mating press,
Pov Doggy style looking,
Pov Doggy style looking away,
Customising prompts
Trained prompts
holding viewer,
Untrained prompts
Creampie,
breast grab:1.3,
butt grab:1.3,
arm grab:1.3,
torso grab:1.3,
head grab:1.3,
leg grab:1.3,
FAQ
Comments (5)
it works great,but it will change the style a little bit
Nice job ! this position was near-impossible to get in vanilla, however there are a few issues :
-style leaks (need higher noising weight and noising bias to focus on composition instead of style)
-saliva goes the wrong direction (upward)
-somewhat overfit (messy hair, long hair, ground, body type and size of both... etc), you need to tag the elements that changes or avoid having more than one or two pictures with same hairs/background/body type...etc.
-ankle-grab is quite hard to get, even more with feet touching the surface.
Thank you! Yeah this one was really hard to generate 1st few pictures for the training data 😅. And also thank you for the feedback I'm still very surface level when it's comes to training LoRAs so it's always good to hear things that I need to improve on.
Will try to fix these problems in the next update (not sure when as I'm currently working on other LoRAs), never noticed the saliva problem so that was a good find will try to train it not to do that if possible.
As for style leaks never been sure how to fix that, do I need to include other styles when training? Or is the noising weight/bias something I can set during training? If so will definitely start playing around with that the next LoRA I train. Thanks again! Hope to get better with each LoRA I make 😁
I've noticed that liquids pouring against gravity seems to be an issue with SDXL in general.
@Oscarc Sorry for the late answer, i missed the notification of your message !
"is the noising weight/bias something I can set during training ?" Yes, by adjusting those you can pretty much set the training sensitivity to different size of groups of pixels, allowing you to either train on the high frequency details (style, or small groups of pixels) or the low frequency details (image composition, or the bigger groups of pixels, which is adapted to your case of a 'position' lora).
"do I need to include other styles when training?" => only if you don't want to fine tune the noising. And if you don't tag every style in the training dataset you have to avoid having a style occuring more than others.
Also, when you tag something in the image, the model understand it's a separate concept than the one you train (like "messy hair") so it's easier to avoid 'forcing' it into every image generated.
FYI i've been able to train really good Loras with carefully tuned noising biases/weights on awful datasets (like nose-hook/triple_nose-hook/mask_nose_hook, the raw datas for this were rare, low resolution, bad quality).
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.