Hi Johan,
Thanks for confirming my conclusions:
- sharpening is not done in the RAW process, but after the image is de-Bayerized
- Non-destructive refers to the fact that the RAW file is not changed
But eventually, the resulting file from the RAW conversion (tif, jpg, psd, etc) is capture sharpened and that conversion is by definition "destructive", in other words the capture sharpening changes the pixels. OK, I'll stop playing with semantics now ;-)
Ah, but semantics can be important!
There is a difference in how sharp the Raw conversion turns out. The Bayer CFA demosaicing is an operation with many trade-offs between artifacts and sharpness (false color and mazing). So one could say hat there is a certain amount of sharpening being done in the Raw conversion itself, in all converters, to varying degrees (sometimes adaptive to image detail).
Given that RGB converted image, there are different approaches possible. The preferred method generally has as much to do with personal preference as with workflow. However, there are objective ways (not always simple ways) of determining the best approach, depending on the goal.
In general one can say that for downsampling, pre-sharpening doesn't help. On the contrary! So if one's goal is to produce a smaller (in Megapixels) image, I'd suggest to leave
all sharpening out of the equation, just go with an unsharpened Raw conversion.
Only(!) if one uses the best algorithms for downsampling, then some (capture) sharpening won't hurt much, if any. Photoshop e.g. currently does
not use the best quality algorithms.
For upsampling/magnification e.g. for large prints, one can apply capture sharpening before interpolation, as long as the sharpening doesn't create visible artifacts(!). If, and how much, one can sharpen prior to magnification, is dependent on too many factors for a general approach. Some algorithms for magnification do better with pre-sharpened image data than others, also depending on image content.
A good series of resampling algorithms is offered by
ImageMagick, and they currently offer some 12+ different algorithms, with additional parameters. Some are better out-of-the-box for downsampling, others for upsampling, some for line-art or text, others for photographic images.
Of course, output sharpening at the final size is (assuming proper execution) always beneficial for the best output quality. Although even then one has to consider things like which spatial frequencies to enhance, and should dithering/noise be added.
Bart