surely this purely has to do with the examples in the training corpus. I reckon SVG will have more illustrative content and as a result a more accurate representation.
I reckon with the right examples in the prompt to take advantage of in-context learning, it could be pretty accurate.