Why do you only use unsharp mask for web only?
Understanding what the two different methods actually do can help with realizing when one or the other will likely be more effective.
Here are three images that simulate the difference between Unsharp Mask and High Pass Sharpening. Consider the top image to be a highly magnified original part of a photograph. The middle image has HP Sharpen applied to the high frequency detail on the right side, and the bottom image has USM applied to the tonal transition on the left side. (In fact, all of this was done by adjusting the contrast of selected areas, but it shows on a large scale almost exactly what sharpening does at the pixel level. The way it shows up here the USM isn't very dramatic, but it does have a little bit of a halo, so it's perhaps realistic.)
First, a "sharpen" tool means a "high pass" algorithm. Some people refer to it also as "convolutional sharpening", which isn't wrong but since convolution can be used for Unsharp Mask too it is a bit ambiguous.
What does "high pass" refer too? The frequency of a signal produced from the pixels (either by different intensity or color) as one scans from pixel to pixel physically along any line in an image. On the right side of those images the dark/white transitions are close together, so they are a higher frequency than the transitions on the left side. Rapid changes of value between close pixels is high frequency detail, slow changes across many pixels would be low frequency detail. A single jump in values is a transient high frequency spike rather than what photographers think of as "detail".
Okay, what a High Pass sharpening tool does is increase the contrast between the highest values and the lowest values of multiple tonal edges that are closely spaced. (And in fact the exact same software can be used, with different parameters, to blur an image by reducing the contrast. This is a reversible process.)
Consider that if you have a 12 MP image that is 4000x3000 there are five times as many data points across a single horizontal line as there will be if that image is scaled down in size to 800x600. If the actual physical display size is the same, the
frequency of the data points is 5 times higher in the 12 MP image.
So consider what happens if you have a 12 MP image and use a sharpen tool on it, and then re-sample it down to a size appropriate for web display! Virtually the entire spectrum of high frequency data that the sharpen tool worked with is
removed by the re-sampling process.
For that reason an High Pass Sharpen tool has little effect on images that are scaled down in size, and virtually no effect if the sharpening is applied before the re-sampling is done. On the other hand, HP Sharpening can be very effective on images that are printed full size and even more so if the image is up-sized before printing.
The Unsharp Mask is different because it affects accutance, or the acuteness of a singe tonal transition. It has an effect that cannot be reversed. And it is unrelated to bandwidth or frequency of the data between spacial boundary beyond a single transition (it relates to high frequency transient response rather than repetitious high frequency data).
If that 12 MP image has a single transition made up of high frequency detail (the transition is fuzzy and several pixels wide) and it is down sized, the detail is replaced with a single transition (If the resulting image has 1/5th as many pixels, a transition that took 10 pixels to go from light to dark will necessarily end up being a transition across only 2 pixels. One pixel is light, the next one is dark). Hence the high frequency detail that an HP sharpen tool might work with is replaced by the kind of transition that USM works on! For that reason USM does little to images that are upsized for printing, and has significant effect on an image after it is downsized for web viewing.
Generally the smaller the image the lower numbers that are suitable for the radius and amount parameters, both for HP sharpen and USM. The exact numbers vary depending on the algorithm used by a given program and also often depend on the exact data too, so it is important to do sharpening
after an image has been scaled to its final size. Also note that for certain data the two methods will actually produce almost identical results.
