I had been working on this for awhile and Astra mentioned briefly the struggles of learning anime art style, in the PONY 7 twitch stream with Ally.
One approach is to use the midas depth map approximation to add a pseudo 3D effect to your flat anime images.
I present a alternate approach which is to use cupy to analyze the image in latent space.
Visually other then blurring it changes the image little, but it drastically changes the latent.
I am still working on this approach and it is highly GPU intensive I think pairing it with the depth map at a random alpha would enhance training considerably without risking loosing the art style.
