>>106012302
Basically the VAE as is does not see these as the same hand. An upright hand is treated completely differently as a hand turned to the right. A large upright hand is treated differently from the same hand but smaller. What EQ does is train the VAE to realize orientation and scale still go to the same hand. My hypothesis is some of the body horror and hand problems we see is actually due to this inherent problem in VAEs and a diffusion model is forced to learn antagonistically that the VAE is "lying" to it about features but it's likely why complex or abnormal poses like a person doing a handstand often results in body horror.