efficientvit
efficientvit copied to clipboard
Why FLUX VAE is hard to optimize?
In DC-AE paper, I observed that the performance using FLUX's VAE was notably inferior. When comparing FLUX VAE and Stable Diffusion 1.5 VAE in my implementation, I found consistent results with the paper - FLUX VAE exhibited significantly slower convergence rates and bad performance compared to SD1.5 VAE.
Has anyone encountered similar issues or can explain the underlying reasons for this performance difference?
channel size maybe?