Trained on 30 images of uncut shlongs (erect, half erect) from different angles and with varying degrees of tip coverage. However, most of the results I get still show the head mostly uncovered. I’m releasing this as a version 0.5 for now because I only did very limited editing of the captions.
Works with men, trans women, futa, and pairs very well with the Loraholic penis sliders.
You can try it with or without the trigger word “un0ck”. It helps to describe if the penis is being seen form above, the side or below.
Validation loss graph:
Description
FAQ
Comments (9)
Nice Lora! I'm curious -- what are you using to visualize the validation loss?
Thank you :)! I`m using AI-Toolkit. This is a relatively new feature. It helps a lot with speed to set the steps for the calculation of it to something higher than 1. It was 20 in this case.
@AugustusLXIII cool, thanks!
@AugustusLXIII looking at the loss graph it looks like you had a pretty good deep trough late-run towards the end, and a notable deep low-loss point between 2750 and 3000 steps. Is this one based off one of those step counts or did you pick another?
@jrewingwannabe947 I picked 3250 based on comparison testing of different prompts. Unfortunately, I can’t avoid all that testing with this graph, but it’s a really good indicator of when you can start doing it. With the gloryhole LoRA for example I got the best results before the graph reached it´s lowest point.
@AugustusLXIII yeah I hear ya. it's like you said a good indicator on where to hunt, but doesn't necessarily know which one is the right one just cuz the math says it should be. you still gotta hunt for the right one w/ prompt testing. I wish they could come up w/ a way to mathematically pin point it - it would save so much time!
@jrewingwannabe947 You could probably create your own test based on some perceptual metric, but yeah nothing is going to fully replace human judgement there.
@WeBeJavn I have a workflow that uses batch analysis for face likeness scoring so finding the best one that reproduces the source dataset, mathematically speaking, is something I have in my current training workflow but it's finding the ones that are the most conformant to the prompt, don't introduce other weird artifacts etc that's the current eyeball-it challenge (with the loss-graph telling me where to hunt). I wouldn't begin to know how to get that prompt adherence or body-horror judgement done through some metric.
Good. It works pretty well with 'diverse penis' lora. Waiting update...