PARTIALLY OUTDATED: WILL BE UPDATED LATER
Blowjobs&Co
UPDATE: The prompting slightly changes, see attached pictured.
It's more consistent from my experiments and also way more powerful and diverse that it was. But it still has issues I want to correct (angle biases, cum, saliva, woman and male genitals). POV can be better on the previous version sometimes, I guess it depends on what you're into.
Warning: This model is still in beta and while it can give beautiful outputs and has a nice flexibility it can also suck sometimes.
All the provided examples are non-upscaled raw gens.
Settings:
steps 20-40
CFG 2.5-5
Samplers: Euler/Simple or Beta, DDEIS/DDIM, Flux realistic
All were done on forge webui (thanks lllyasviel)
The objective of the model is to reproduce with a high fidelity and a good flexibility: handjobs, blowjobs and its derivates (deepthroats, cumshots etc).
POV blowjob/sucking the tip of the penis/sucking an item pictures are easy to get right.
POV handjobs works most of the time.
Distant blowjobs works but inconsistently and requires a precise prompting (a character is kneeling and performing a blowjob on a standing man, then describe the character, see examples)
For the rest I let you expirement and give me some feedback.
Description
First iteration
FAQ
Comments (6)
it can also suck sometimes!!!! LOL
works very well ! The distance shots are a bit tricky to get to work. Maybe a Lora that is just for that so it doesn't get confused with all the other concepts with similar wording?
I'm not really interested in doing that, my main goal is improving my training technique. I want it to have a good generalization, as I said, it's a beta. I mostly know what is wrong on this part in the tagging and I think I found a workaround for the next iteration.
That said it's most of the time fixable with a prompt that pleases its bad biases. Thanks for the feedback.
OK so I've split the dataset to try to fix the remaining issues, I'll publish the intermediates POVOnly/DistantBJs for those interested.
@soKyra ah cool! you see improvements with the split loras?
@WhatTheGuy Tbh it's worse than my last training on a full dataset (unpublished) because I removed the anatomy and so it lost a bit there without being better, that means the rest of my dataset is quite stable. But it did help me spot some important issues that were hard to guess where they were coming from before the split. Upcoming models should be much better.