The difficulty is "prove you did it right". I'd guess that only works for cheap-to-verify (trapdoor"?) functions. However, Intel SGX does enable this. You would need an SGX-attested renderer, but the output would include a signature on the inputs in addition to the actual output image. You could verify this signature was made by the SGX-secured renderer code, then you'd know it executed as designed (i.e. the person didn't just draw an image in paint).

SGX would also enable hidden contract data. Right now, you can't publish a contract like "guess my weight and win $100", because the guess code is public (has to be, for miners to exec it). With an SGX-enabled system, this is now totally doable. (Though, you must trust Intel, and assume hackers can't reverse engineer the Intel chips with magic physics or other breaks. So it's probably safe for a few millions but not billions of dollars. (Upper bound, for a few hundred billion you could acquire Intel and compromise SGX secrets.))

SGX is pretty exciting and I'm in the early stages of getting at least one project working with it. There's also research called "Town Crier" to enable TLS-retrieved website data into Ethereum contracts.

For all the people that cry about SGX being big, bad, DRM, it really has its upsides. Too bad Intel requires special licensing to use it - it should just be enabled for all purposes. Maybe they're worried about malware backlash?

Also, an SGX-enabled Ethereum, apart from the Intel/SGX trust issues, means you can skip the computation and just verify the result. It also means you could replace the proof-of-work multi-hash with a single iteration plus a sleep.