Setting alpha

[mashrurmorshed]

In the paper there is a hyperparameter $\alpha$ that is used to scale norm_grad. Usually it seems to be set like alpha = 1 / norm or alpha = 0.1 / norm (mentioned in Appendix C in the paper, in each task subsection). However, in the code it seems there is no such scaling done? Why is that the case / what difference does that make?

For example, for inpainting, the paper mentions $\alpha = 1 / ||y - P\hat{x}_0||$. Doesn't $ ||y - P\hat{x}_0||$ get smaller with increasing T? That would mean the gradient constraint gets stronger the closer you are at the end of sampling, which is not what I would usually expect.