Post

I was kind of blown away by an early synthetic training data system that I saw at GTC2018 where they used a 3d asset generator + a style transformer to make a real time driving environment for testing autonomous cars virtually

SE Gyges@segyges.bsky.social

the last ~year or so of image models are actually black magic wrt things like this, and from this you can infer that they have to have very good training datasets encompassing this sort of request

very, very rudimentary but since their dataset was trained exclusively from images produced by the real cameras on board the actual cars, the result was nearly identical. They just had a game engine running underneath and people could make new test courses

0 reposts4 likes

No replies yet.