Why you use diffsusion's q_sample instead of adding noise directly to the input? (DRM.py)

[ChoiDae1]

Hello!
I have one question about your implemtation code.

In DRM.py, I understand DiffusionRobustModel's denoise function plays the role of both noising input image (by using predicted timestep) and denoising.

But, in general certify scheme, input image is first augmented by directly adding noise and then passed classifier.
So, I think below procedure is more correct.

Given sigma,
add noise to input image by directly using this sigma   -> denoisng this input by using predicted timestep

In your implementation, noise-level is influenced by predicted time step. This means you use estimated sigma(not given sigma) for noising input.

If not, please explain about this.
Thank you