The noise image obtained by inverse DDIM is not like a Gaussian distribution!

[whisney]

I try to use DDIB to achieve the MR-CT image translation. I have trained diffusion models for unconditional generation of MR and CT images respectively. Given an MR image, I use MR diffusion model and inverse DDIM to obtain the corresponding noise image in latent space (steps=1000, the total step of trained diffusion model is also 1000). However, this noise image does not seem like an isotropic noise distribution (as shown in Fig. 1 in the paper):

Taking the obtained noise image as the input of CT diffusion model and sampling with forward DDIM, the generated CT image is not ideal.

[Bayern4ever-dot]

Could you please share your procedure of training on your own datasets?I would appreciate it if you do so!

[xiayhh]

I would also appreciate it if you could share a little about your own training process.

[hello-world-hhh]

I also encountered the same issue