I didn't try any LORA training since I wanted to make the hypernetworks work the best I can get them to work, hoping that more parameters > better quality. So I don't know if this is better than a good LORA or not - would appreciate feedback. After using it, I find it very imperfect and not exactly what I wanted. With some lucky prompts and seeds it is exactly what I wanted and I think you could make some pictures that are indistinguishable from ones that were made by the original artist. But when you aren't lucky the quality can be horrendous. Most of the gens I did were at low cfg (2 to 3) and full strength of the network so that may be part of the reason. Check grid png in files to see how cfg and strength affect the pictures.
2 hypernet files that are probably not that different in output, pt3 version was trained only on "art by pochi_goya" tag and 4-l-500721 version was on deepdanbooru tags I also checked by hand + "art by pochi_goya" at the end. Probably makes sense to end prompts with art by pochi_goya.
MODELS USE DIFFERENT NUMBER OF ACTIVATION AND NORM LAYERS THAN DEFAULT. Activation functions and norm layers end before the first to last layer. Also norm layers don't use elementwise affine. If you use a1111 then you can use the hypernetwork.py file.
I could write another long github blogpost on everything I found doing this but it makes no sense if this is worse than a LORA so please do tell if you find it unusuable or noticeably worse.
