VP (variance preserving) vs VE (variance exploading)

[wangdelp]

Seems to me the add_noise function in some schedulers are VP, and some are VE:
VE example: where in noisy_samples = original_samples + noise * sigma, the original_samples is not scaled to be smaller, so variance is exploding.

diffusers/src/diffusers/schedulers/scheduling_dpmsolver_sde.py

Line 527 in 83bc6c9

def add_noise(

VP example: https://github.com/huggingface/diffusers/blob/cf6e0407e051467b480830d3ed97d2873b5019d3/src/diffusers/schedulers/scheduling_lcm.py#L594C19-L594C79

I wonder why that's the case?