Is upsizing, sharpen, downsizing superior to just doing the deconvolution in the raw processor?
It's hard to say in general, it depends on the source file and subject matter. However, when you run into problems with direct deconvolution, you may be able to still do something with the round-trip approach. Files from cameras without AA-filter may be problematic to sharpen, due to the aliasing that's present in the plane of optimal focus.
I have 3 versions attached. This is from a 100% crop of a snow bank evenly littered with small leaves. I picked a crop spot with sharp focus at the bottom moving to de-focused at the top.
All are converted in Raw Therapee to 300% output then downsampled using bicubic in images plus.
I assume that is with sharpening at the up-sampled stage?
Its hard to say which is better all around. What do others think?
On the sharpened image I see a reduction of some low contrast micro-detail and something that looks like a dark halo. Don't know what is causing that, but perhaps a somewhat lower smoothing, and different radius (seems a bit small, but then I have not measured your actual system PSF) will change that. An attempt with FocusMagic on your unsharpened (and mildly deconvolved) version gives quite a different result, retaining low contrast micro-detail while increased resolution.
FocusMagic usually gives very good results at 100% size already, but one can combine it with a Luminosity Blend-if layer which tends to make an even more subtle effect, tends to reduce noise, and avoids clipping issues, with the added possibility to use masks.
My alternative solution offers a possibility to apply deconvolution sharpening on up-sampled data, preferably without previous sharpening, so it can be used as capture sharpening at a sub-pixel level. The resampling algorithms should preferably be very good with minimal artifact creation. The benefit of using this method in combination with FocusMagic is that we because we cannot enter know blur radius information (or if we do not know it), we can use visual feedback in the small preview to hone in on the sub-pixel optimal level that matches the FocusMagic algorithms used.