Lpixel loss

Hello, I have read ：Diffusion Denoising Approach for Monocular Depth Estimation。 This is a great job, but there are a few issues:

Firstly, does Lpixel loss only supervise The depth decoder and encoder? If so, it seems to contradict the later Self Diffusion because Self Diffusion did not use real depth labels, so this training has not used real depth labels.

Secondly, if denoising is not performed in Depth Latent Space. Do we still need to keep the Lpixel loss ?