Difference between LDM and pixel-level DDPM.

[Kenny-K]

In pixel-level DDPM, the linear noise schedule is established with beta_0 = 1e-4 and beta_T = 0.02. However, in LDM, the range of (beta_0, beta_T) is reduced. For instance:

params:  
    linear_start: 0.0015  
    linear_end: 0.0195  

Additionally, DDPM learns a weighting on the variance, while LDM fixes the variance as posterior and eliminates the VLB term.

What is the rationale for this discrepancy?

[tusharkumar91]

Hi @Kenny-K , were you able to find some information regarding the reduced range of betas ? I was having the same doubt and wanted to get some clarity on reasons for using this schedule