Comment on Explaining the SDXL Latent SpaceparentComments−shenberg2yI suspect that weight initializations are geared towards inputs being normal random variables with mean 0 and variance 1. Deviating from that makes the learning process unhappy.
Comments
I suspect that weight initializations are geared towards inputs being normal random variables with mean 0 and variance 1. Deviating from that makes the learning process unhappy.