v2.5 - thanks to the comment by a fellow user i was able to find a good setting for training that not only allowed for more controlled generation but slightly better results with other models, clearly i still got a bit of testing to do before its working, but i expect to be increasing the dataset for v3 to get even better results overall.
-(human) in the negatives might help with the human looking torsos,
-dog in the negatives might get rid of the dog looking snout if you prefer not to have it.
-"a dragon holding a knife" or similar captioning might work along with the usual trigger words and tags
-still works best using revanimated, mistoon, dreamshaper, etc
-playing with weights helps, some models require more and some less.
updated to allow for more customization and it should work better on a few more models although it all depends on their kobold training. also might lose some control so playing with weights is recommended
- after some testing, v2 uses dragon as a trigger word along with kobold,
v1-
kobold, snout, horns, claws, tail, scales, white background, yellow eyes, fangs, red skinKobold race from Dungeons and Dragon, trained for 10 epochs,
Some models are hit or miss, seems to work good on revanimated, some cartoon or anime models, and sometimes dreamshaper
personal favorite epoch is 9 so ill leave that one for download.
the rest will be uploaded on huggingface along with a list of weight distribution.
https://huggingface.co/BakingBeans/KOBOLDS_DND/tree/main/checkpoints
*v2 epochs will be available soon.*
Description
9th epoch of the training
FAQ
Comments (6)
Fantastic stuff, I've been struggling to get a lot of the monster races to prompt well.
Generations from V2 look great! Any additional info worth sharing about V2? (Suggestions for constraining for differently colored kobolds? etc)
Do more races please. You have my attention <3
Most of the images look a fair bit overfitted. Have you tried using Min SNR Gamma = 5 while training?
It essentially doubles to pentuples the amount of epochs you can run, before the visible overfitting happens.
Very good goood
This is exactly what I need. Why is it archived? =[






