Just from from some quick tests, I'm not fully satisfied with V2. It is ok and an improvement over V1, but I am going to make some changes for a V3 to see if I can improve the finer details. I'll update this post with the 3rd version at some point. Still, V1 and 2 aren't complete garbage and you could probably make some decent images with them!Update (1-29-25): I released a 2nd version with more training steps and double the dim/alpha. I'd say it's a noticeable improvement, but I plan to re-evaluate the dataset in the future to see what tweaks can be made, images to remove and add, etc. Still, V2 is at least more decent now than V1.
