ROWBench: Do Video Models Render What the Program Specifies?
PROWBench tests whether generated video follows the exact events recorded by executable world programs.
The paper introduces 170 constructed episodes and 600 proxy videos with logged entity states and timestamped events. Those records can include events outside the camera view, then render synchronized views for checking visible consequences. The benchmark covers first- and third-person perspectives, with some multi-view observations. It evaluates entity control, long-horizon memory, and alignment between the prescribed timeline and what appears in the generated video. HF Daily Papers' note
The paper introduces 170 constructed episodes and 600 proxy videos with logged entity states and timestamped events. Those records can include events outside the camera view, then render synchronized views for checking visible consequences. The benchmark covers first- and third-person perspectives, with some multi-view observations. It evaluates entity control, long-horizon memory, and alignment between the prescribed timeline and what appears in the generated video. HF Daily Papers' note
score 5