In the above slide, I understood that the latent diffusion model operates on a vector-quantized latent space.
Then, I wonder about how we can define the forward (diffusion) and backward (denoising) processes in this discrete latent space.
And also, what are the implications of using a discrete latent space for LDMs? Is it just a form of regularization, as mentioned in the slide?
In the above slide, I understood that the latent diffusion model operates on a vector-quantized latent space.
Then, I wonder about how we can define the forward (diffusion) and backward (denoising) processes in this discrete latent space.
And also, what are the implications of using a discrete latent space for LDMs? Is it just a form of regularization, as mentioned in the slide?