This LoRA allows depicting a person smelling / sniffing their own armpit.
Surprisingly Pony doesn't already know this concept (at least with tags I tried), so I decided on creating a LoRA for it.
Main trigger: sniffing armpit
Additional tags (ordered by tag frequency): exposed armpit, clothed armpit, arm lowered
(the last one had so few images it is really hit and miss; it is better to also have exposed armpit in the negative for clothed armpit, as the term armpit already makes pony quite happy to generate ... well armpits)
Suggested LoRA weight: Depending on the style you want 0.4 – 1.0.
And the explanation of pony being so bad with it is probably that there are barely any images tagged accordingly on different boorus.
And with that we can talk a bit about the
Training
Specifically, I collected 36 samples from different boorus (not really cherry picking, they were all good ones I could find).
I then generated 170 additional images with Pony Diffusion using Control Net (mixture of Depth and Pose models). For that I used random art styles, gender, ... .
Input images for the Control Net were both drawings and real photographs (unbelievably many stock photos of this exist). They were sourced from conventional image search and 19 of them were added to the training images.
This resulted in 225 training images.
All of them were then tagged using wd-swinv2-tagger-v3 by SmilingWolf and afterwards the 4 tags listed above were manually added.
Afterwards I added masks to the images using first RemBG (Human) and then ClipSeg for the text Arm, Armpit, Face. As the dataset was small and a quick glance showed that not all masks were correct I also did a quick manual pass correcting masks.
I then trained the LoRA using OneTrainer.
The relevant training parameters were:
Prodigy optimizer
24 epochs @ 560 steps
10 image repetitions (with image and caption variations)
Batch size 4
Using image masks, with unmasked probability of 0.03, unmasked weight 0.02
1024 resolution with aspect bucketing
LoRA rank 48, alpha 2 (later resized to target 32, with sv_fro 0.99)
Training was done for about 8 hours on a RTX 4090.
If you have any additional questions feel free to ask.
Description
FAQ
Comments (4)
Gross. I love it!
Thank you for your service to all the musk degenerates like me, especially for furry males, who went crazy realizing that not one single model (ive found) knew even remotely about this concept. The closest ive ever come is getting another char to sniff the other chars pit, and that was very likely only looking like that, on accident.
Edit: I have tired this now with multiple models, and sadly i have a few problems. I only tried making furry males so far. When using it, even on low numbers, it will force a style onto the image, like the one shown in the preview images here, even if i use another style, or even other style loras. It will also often delete many details or most of the background. On low numbers like 0.4, i can almost never get this to actually change a characters pose, like turning their head to their pit, but the effects described above are a bit weaker. On higher values like 0.9, it does often work, but the other effects ruin the image for me. Maybe i use too many other prompts or want a too complicated scene, most images here using this dont have as many prompts and concentrate more on just using this. Models i used were steinxl_B4, nemomixXL, autismmixSDXL, all pony merges.
The training also included multiple sample images of furries (it was the source where I saw most of this concept). Also the training data didn't really have one specific style. If you are copying the provided prompts one to one, you have to watch out you do not copy style tags. Pony has some inherent styles (which are trained artist styles without training for their names). For main preview image for example the tag "aqe" will trigger that style and its bias will also be quite strong.
If that is not the problem, I don't really know what the issue could be. (Or if you use very diverse prompts, the concept might just not help as much, as the training data was not similar enough).
@DiffusedIdentity Ill write you a dm with some pics ive made to show the effects i mean, cause i cant do that here.