🐛[BUG]: "stochastic sampler" is hard coded as deterministic within Corrdiff?

[joshdorrington]

Version

1.2.0

On which installation method(s) does this occur?

No response

Describe the issue

Hi I am working with the corrdiff model, which calls the stochastic_sampler class to generate samples for its residual net, within its generate.py file. When I look at physicsnemo/blob/main/physicsnemo/utils/diffusion/stochastic_sampler.py I see that the stochastic terms in the sample generation are turned off by default (S_churn=0). The corrdiff code doesn't pass any config kwargs into stochastic_sampler that can turn these on, so the model looks to be hardcoded to deterministic sampling.

I see that the corrdiff/generate.py uses an array of different random seeds to obtain an ensemble of samples from this generator, but this is not a stochastic sampler in the sense of Karras 2022's EDM, which the corrdiff code is otherwise following closely.

Is this a bug? Or is it an intentional model choice? If the latter, could someone please clarify the motivation for the choice? :)

Minimum reproducible example

Relevant log output

Environment details