Have a look at Non-Uniform-Sampling

Towards the end of the NUS presentation here shows NUS is ideal for HSQC because there is only one y point for each x point, but as the number of points in the y axis increases, so does the requirement to have more increments to sample the Y dimension. So if you are interested in the main 2-3 correlations in an HMBC or a COSY, the default 25% of 256 will work, but its reliability will suffer more if you may be interested in more peaks, so you would need to sample more increments.

The situation in the NOESY is that the main detected peaks could be the diagonal peak along with its truncation artifacts, so without large sampling, combining NUS with NOESY could be unreliable.

Regarding which sampling, just use the default Poisson-gap. Unless you have knowledge of the T2s for each peak, there is no point in optimizing this.

For processing I typically use hmsIST: