First up, SFML wasn't developed as a rendering system for GUIs, thus if your aim is a GUI-only application, then you maybe should consider other possibilities.
- Render every GUI element each frame.
- Render only when an element requests un update, on a screen-sized texture that is rendered on the screen every frame.
There isn't
that much difference, because for both operations you'll draw all the pixel of the whole window, thus what would count is rather the logic on calculating the positions and perspectives etc., which if done right shouldn't take up that much time.
Knowing the developers of
SFGUI, I can only tell you that the bigger performance breaker is the rendering part with SFML.
TBH I don't really see any reason for the project. Everything that has to do with web frontend development are very often just 'hacks', because you're very limited on what you can and can't do. If you have C++ and SFML (/OpenGL) on the other hand you get (nearly) maximum freedom and thus trying to copying the hacks & workarounds for a limited system into a freedom system seems kind of strange.
