>>68203013Last one for this thread probably. G'night /vtai/.
>>68206831Model wise, it's simple merges, together with a few Loras, however the VAE is completely Animagine, and the CLIP is 100% Animagine sent through some of the Loras to add their CLIP weights. I had to use the Animagine VAE, as the pony Vae just completely mangled the image, even if the models were merged 50/50. And the CLIP I used, as otherwise it forgets what the characters/chuubas are.
So, all in all: Simple model merge, with Animagine VAE and (slightly altered)CLIP.
Yeah, it's rumored that that's behind GPT nowadays.
