Diffusion models make pictures from noise but often miss what people actually want in the prompt or what looks good to humans.
This paper teaches video-making AIs to follow real-world physics better without retraining them.