Žiga Kovačič releases MPMWorlds to benchmark how well vision-language and diffusion models predict physical dynamics from video · Digg