v5 Use weight at 0.92

works on its own.

Read the "about this version" description

Everything below here is outdated.

v4 NEEDS v3.5
Use v3.5 weight at 0.9 and v4 at 0.93

v4 is like an enhancer and v3.5 is the base. They work well together. But you can use v4 together with other male anatomy LoRas as well.
Read the "about this version" description

I gave up on merging, so I'm releasing both versions to be used together.

I recommend using a penis LoRa to get more details. Xpenis Uncut by @wolffur666456 works great here at 0.4-0.7. I still have to test his ass LoRa interactions with mine.

No trigger word, just reference a man from behind or showing ass

Some supporting trigger words used on the dataset:

glans, large penis, small penis, medium-size penis, wide open anus gap, closed tight anus gap, penis upside-down, scrotum, testicles, low-hanging testicles, hairy, shaved 

V4 learned a great deal of details and anatomy, but it is not good on its own. IMO this is way better at prompt adherence, but one is not good without the other.

I did a block weight analysis on both versions and tested its interactions. I rescaled the rank of both LoRas and I think there were no expressive quality hits. I sometimes preferred this Rank24 LoRas so I decided to go with them to spare everyone HD space.

V4 is an insane version made with too many continuations. I trained v4 a,b,c,d,e,f,g and finally this OC which was supposed to mean Over Cooked because I upped the LR so high, and it still didn't burn. To be honest, I think kohya bugged with this continuations LR adjustments, it probably was all using the firs LR that was super, super low.

This also had regularization images, with probably dragged this down. I currently don't think this is good with kohya. Only with AiToolkit with reg_weight at 0.25.

I have plans to train a v5 to be good on it's own.

Everything below here is outdated. The table is outdated, I might fix this later. Or not.

v3 use weight 0.68

Dataset update for v4 05/09/24:

I've been focused on regularization and recently on my bounding box program, it's now finished. I'll keep up with some more manual captions and soon try a v4. I pruned the dataset a little bit.

v3 use weight 0.68

Read the "about this version" description

v2 USE WEIGHT 0.53

Works better with Distilled CFG Scale of 5 or above and 50-60 steps (yes it needs a lot of steps)

No trigger word, just reference a man from behind or showing ass

it responds quite well to negatives (use the dynamic thresholding CFG-fix),

This is not great yet.

This is my first step going away from Bussy. This will no longer use that trigger word. I want this to be not just a strict pose from behind but more like full male asshole Lora from any position. I want to include more action like insertion, rimming and maybe anal sex.

But for that I need to increase the dataset and improve captions. This is a first try using the same dataset of Bussy (without any trigger word). It needs a whole redo with better captions and less danbooru tags.

The dataset has 10% of color associated regions and it was enough for flux to understand most of the position of the concepts (ass, anus gap, penis, glans and scrotum). But the quality is really bad for now for v1. v2 is a little better.

I hope with better dataset and maybe better trainers and training parameters it will be better in the following versions.

Problems:

• bad at close-ups

• can't do hairy well

• penis is really weird shaped most of the time, v2 is better but still bad

• without negatives, it's hard to control penis size and visibility

• overall looks low-res, v2 is a little better

• v3 text habilty suffers, will be better in future v4 with regularzation images

Trained with Ostris ai toolkit.

for a better description, read SDXL Bussy Lora page.