A question about the DDIM loss.

[ganyz]

Dear author: Firstly nice work and thanks for code! A question about the ddim loss as shown in the fig.

I don't know why , in the forward process (adding noise), it adopts the blur_depth_t (the x0 result of xt->x_t1->..->x0) instead of the gt_depth?

[duanyiqun]

hi there, the name blur depth_t here is just a name pattern.

Please track the blur_depth transition. It should be refined depth latent. As we mentioned in the paper, gt_depth might be sparse in outdoor scenarios, so we use a self-diffusion formation (diffusion on refine-depth from itself).