Support for REPA-E vae?

[mozophe]

REPA-E vaes (https://end2end-diffusion.github.io/repa-e-t2i/) released recently are expected to significantly reduce training duration. Would just changing the vae to a repa-e vae work?

I am not entirely sure on how to use this for training.

There is also a comment by REPA-E's team that existing qwen models won't work with REPA-E: https://huggingface.co/REPA-E/e2e-qwenimage-vae/discussions/2#68f94dba387151a132faaf6c