Okay, I said 3? 4? Drawings? I probably meant about 9 X_X. EEP! But... I think I finally have my drawing set for V2 of my Dante Lora! Painstakingly crafted and prepared in a way that I hope the AI will be able to easily consume and learn from! It's a total of 38 pictures! The previous set only had 37 base images to go off of!
. . .
One more picture? I thought you drew 9 of these things, Dante! Actually I drew a whole lot more than 9... The new data set replaces a ton of the old data in ways that should dramatically improve the AI's learning capabilities (in theory). In fact, there isn't a single image in the new image set that was in the old one. At the very least, every image was updated to add shadows and some details to bring it up to spec with the new data set. The old data set had about 8% assisted images, while the new set has 8% non-assisted images... That means that 35 of the 38 images were AI assisted and all 35 of those images were not in the previous training set.
The new data set has fairly consistent anatomy throughout, about 15% have 4 fingers instead of 5 (that will come back to bite me) but that's far better than my 50% mix I used to have. Likewise, number of toes is locked in at 3 across 100% of all images and ears dashes are present across all images. The new set also uses a lot of full body shots that give excellent views of Dante's entire body, and they're all high resolution so I can often train a full-body, half-body and headshot from the same image, increasing the value of each image so that I will get way more power out of these 38 images than I could out of the previous 37. They are all fairly consistent in their anime drawing style as well, so the AI should have a clearly defined target to hit.
I also have all files locked and loaded for changing clothing colors, or to swap Dante from one background to another. This way the AI learns Dante "can wear clothes" (though I don't blame it if it concludes that the panda fox is perma-nakie) but I don't want it to learn he only wears a specific color. Likewise, I sometimes would find my other LoRA was occasionally overtrained, especially on
So if I ask it to have Dante jump or mentioned lakes or forests at a higher LoRA settings of about 0.6, it would often copy the pose and background. That picture doesn't even exist in the new set, and hopefully the new changes will help the AI get less attached to a specific picture in the data.
Ultimately, I'm super stoked about what happens next. Putting some Dante's on backgrounds, maybe adding in shadows and cleaning things up. Write up some prompts for close to about 300 pictures X_X and then prepare the LoRA to bake. There is a new version of YiffyMix out, too, which I haven't had much luck with, but maybe I'm doing things wrong and it's actually super amazing. I'm also planning a few extra experiments with this new LoRA set which are now possible. If any of these work out, my results from my future artworks will be a huge leap forward :D!
