The discussion in this thread is of great interest to me, as I am now working through a similar process to develop a world using Wilbur teamed with Photoshop.
What I've found is that perhaps unsurprisingly the higher the resolution you work with, the better results you get. I started with a test model of 6000 x 3000 representing the entire world, and ran it through precipiton erosion and incise flow procedures until it started to look about right. From the start this was just a test, but it really brought home to me the "scale" of the erosion actions, and the fact that I would indeed need to go much higher than that resolution to get a world that actually looked like a world instead of just a region.
Next I moved up to a 21600 x 10800 model, working this one through to close to finished too. The increase in resolution posed a number of problems, not just because of the time it took to complete an action but also because of the sheer scale. I found that mountains and hills worked out well - far better than at the lower resolution, although still too big to make a realistic world. But the plains were featureless expanses of nothingness. Lessons learned, my next model starts out with much more texture in the plains for the erosion to wear down but still leave something interesting behind.
But before working through the next world model, still limited to 21600 x 10800 for practical reasons, I decided to zoom in on the most important region and develop it separately. Incidentally, I reprojected it into a stereographic projection so that the shapes would be true (or at least truer), because the latitudes involved are distorted in the usual equirectangular projection. For reasons I don't understand, I was unable to get the file working right at the 14404 x 11492 resolution I started out with - fill basins was acting very strangely and corrupting the map. Cutting out the surrounding sea until it got to 9998 x 9999 solved the problem. The area depicted is about 5000 km across, and my starting design had mountains, hills, and textured plains to work with. It went rather well.
All my tests have reinforced the idea that working on smaller regions at high resolution is the answer. This seems to agree with Waldronate's comments in this thread, which is very reassuring. For me there are two things to bring away from this: for my purposes, it may not be necessary to get down to real scale of 25000 pixels to 1000 km; and it is probably easier and more practical (or in fact necessary) to work on the world piece by piece as necessary rather than trying to do everything in one fell swoop.
On the issue of scale, it seems to me that with a fantasy world it is acceptable - indeed, it's standard practice - to start with a reasonably rough map and slowly add detail as things develop. It's nice to have as detailed a base to work from as possible, of course, but it's just not practical to do everything in high resolution from the start.
At the same time, working with Photoshop allows height maps to be worked on in smaller chunks and recomposited back into the master map. This is absolutely necessary in the case of the poles, which really need to be done in a separate projection from the rest of the map. I use Manifold to accomplish this; I started out with G.Projector, but it doesn't work with high resolution images and is less able to round trip an image (e.g. from equirectangular to stereographic then back to equirectangular). My current plan is to work on slightly overlapping areas and slowly build up higher resolution maps of the world. Eventually these can be composited back into the world map - or not. I haven't decided yet.