noisy_x scaling before diffusion models

[florinshen]

Hi, I noticed you apply a normalization with on noisy rendered images with a constant (1 + sigma^2). (https://github.com/pals-ttic/sjc/blob/main/adapt_sd.py#L133).

I am confused with the intuition to scale the noisy data with such constant. Is this scaling to keep the deviation of input at 1 ?

[w-hc]

yes and that's how diffusion models are trained in general
see https://arxiv.org/abs/2206.00364 table 1