This is an experiment: I drew a minimalist dataset of icons in gimp and refined with the output of a first model.

Trained with basesd, tested with dreamshaper.