For this experiment I staged a short downhill longboard sequence in Blender, then tested seven AI video workflows against it. I cared about the performance as much as the image: the turn, the hand reaching the road and the final crouch had to survive.
Some attractive outputs changed those actions. The breakdown compares the results with the previz and shows the input repairs, so it is easier to see which decisions stayed under control. There is also a separate comparison of four upscalers.