The latent can be upscaled directly but the loss in quality is important. That means that the second pass after the upscale has to be done at a very high denoise (0.55+).

This option is offered as it's very inexpensive computationally and it can be useful if you don't need the final image to be 100% of the initial generation.

If the final image can be rather different from the first pass, you should use latent upscale as it's faster and very light computationally.