sayakpaul opened this issue 2 years ago · 0 comments
Since you mentioned using self-conditioning from Bit Diffusion here, maybe it also makes sense to use asymmetric timesteps as explored in the Bit Diffusion paper?