We currently using "clipboard mode", instead of "parsoid mode" for rendering the visual diff, as the former produces full view HTML, however it also generates (at some expense) data model information that is not really required for view-only modes (such as the visual diff, or previews), in particular reference nodes can be particular slow, especially when the converter is invoked multiple times, as is the case in the visual diff.
We should introduce a "preview mode" to the converter so that nodes can quickly render view-only HTML. This would also allow us to fix the issue with some nodes not rendering anything meaningful (T170134/T170107).