Why a ?????? LoRA?
I've noticed that there are not a lot of AB/DL related models or LoRA, and those that do exist tend to focus on female content. I wanted something that could be used on a variety of models for different results that was also capable of generating male content without having to filter heavily to de-feminize the results.
Testing Results
v0.6:
This version has been tested on a variety of 1.5 derived models with a fairly high rate of success.
It tends to yield better results with a CFG of 6 to 7 when used on the runwayml 1.5 model but has shown decent results anywhere from 6 to 9 on other models such as
It is much better at playing nice with other LoRA than my previous version and it still does well with key adjustment LoRAs like
Version 0.6 has been redesigned from the ground up using the same source image set as 0.3 but with better categorization and keywords to allow it to specify the types of ?????? you want to use (i.e.: ?????? vs pullup) and, while not 100%, does tend to reliably produce the requested type for the most part. It does help sometimes to reinforce the prompts like so, where it does understand the first prompt generally but the abdiaper, wetdiaper, and pupdiaper keywords help reinforce it if needed:
wearing a ??????, abdiaper
wearing a wet ??????, wetdiaper
wearing a pullup, pupdiaper
v0.3:
This version has been tested on a variety of 1.5 derived models with a fairly high rate of success.
It tends to yield better results with a CFG of 8 or 9 when used on the runwayml 1.5 model but has shown decent results anywhere from 6 to 9 on other models such as
It plays OK with some other LoRA but is a bit over trained so results vary depending on what you're trying to mix it with. Though it does well with key adjustment LoRAs like
The Final Word
While I have no interest in monetization, and this is purely a hobby/passion project for me, I would greatly appreciate it if these LoRA aren't abused for financial gain or misused for illegal content.
While my testing has found that it can work for most age/gender/nationality combinations, your results may vary depending on your positive/negative prompts and expectations.
I have found, with both versions, that some types of non-photorealistic styles can work with some creative prompts and weighting, but it was primarily trained for photorealistic results and I make no guarantees of success outside of that.
If I can find a decent subset of images to work with I may still add to the possible prompting capabilities on future versions if the time and interest are there (such as ?????? under a skirt).
I welcome reviews and feedback and I hope this LoRA is useful for those who might want to use it.
