Question about adding noise to input

[baneitixiaomai]

Hi!
I am currently studying your NCSN project, and there is a bit confused about the handling of the input:
on Line 136 in the file ncsn/runners/anneal_runner.py --- "X = X / 256. * 255. + torch.rand_like(X) / 256." Why add a small Gaussian noise to X，and the predefined Gaussian noise is added in function "anneal_dsm_score_estimation"
I am curious about why this is necessary and would appreciate any insights you could provide.

[linzangsc]

same question. and I observe that this simple turbulance will affect the convergence of loss. does anyone have any clue?