trained on kakure_eria's work, combines particularly well with my other loras, if there's demand for it, I'll train a version better at differentiating the different views.
Description
FAQ
Comments (5)
the most distinct part of this style is the veiny penis and the thick outlines. Works well with my other loras.
If there's demand for it, I might retrain a bit more thoroughly. Improvements can be made in how certain poses are handled, namely from_side and between_legs needs to have the repetitions changed.
How does it work? I mean I though it was just a style, but it seems this lora also brings some positions my checkpoint doesn't have? Like I tried your prompts and they are quite the same in terms of end result, otherwise without lora it won't be that good. So originally I though style loras just a reskin for your checkpoint, but also bring something like poses, objects?
How the style functions depends on how it's trained. You can get style loras like what you expected, ones that function similar to a skin, without preserving the concepts of the source, and sometimes that's the best way to do it, like when there's only a few images. In that case, that's really your only option.
In this case, the artist has a large and diverse body of work, so I trained on a a few hundred very different images, I made sure everything was tagged decently, and trained for about 60k total steps, so a lot of the poses and concepts of the source were learned by the model, but specifically in the artist's style.
This is even more pronounced with artists that have a heavy focus, like mizumizuni; if you trained on their work, it'd basically be impossible not to also accidentally train the model to understand deepthroat.
Based. There's another Kakure Eria model floating around as well, I've had some good results combining the two together at low weights.
How many images were in the training set? Which SD checkpoint was it trained on?