I don't know why the plate solving would be more reliable. Perhaps you're using the accurate scale now and you weren't before?

Re the number of images the overlays are kept in memory. (A less lazy coder than me might have done some on-demand loading, perhaps I'll do that in the future someday, but probably not really soon.) So, whatever virtual memory the images take will be your limit. You could experiment by adding 10 more and seeing the increase in virtual memory for the kstars process. Then add 10 more and verify you have a similar increase. Then you could decide--let us know.

Note: The file size won't tell you the answer--jpeg compression won't be taken advantage of. The full image stored in memory in a QImage (e.g. so the memory use will likely be something like width * height * bytes-per-channel * # channels. That is, if all your images are RGB and the same resolution then they'd all take the same amount of memory. If you're curious, you can see the code that loads the images here: