Skip to content

Comment on Explaining the SDXL Latent Spaceparent

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.

AboutSource Built by g1lg1l

Hackerly is an independent reader for Hacker News, built on the public HN API. Not affiliated with Y Combinator.