This is pretty cool. I wonder what the training dataset was. I assume some kind of mapping between 3D model data and text descriptions? There can't be very muc

This is pretty cool. I wonder what the training dataset was. I assume some kind of mapping between 3D model data and text descriptions? There can't be very much of that data.