This benchmark ranks world and video generation models on whether their videos pass for real camera footage, judged blind against each other and against the real clip each scene was taken from. We took 25 real clips of everyday motion, from a plane landing to coffee being poured, and had every model generate four videos per clip from its opening frame and a one-line description. The rankings are updated continuously from 618,210 responses across 85,098 pairwise human-preference matches. All data and judgments are public; see the methodology for details on evaluation and scoring.
Table 1. Overall standings, World Models Realism benchmark. Scores from a Bradley–Terry fit rescaled to Elo; intervals are 95% Wald intervals. Click any row to view sample outputs in Figure 1. The real footage is scored alongside the models as a reference and takes no rank.
Figure 2. Overall Elo of each model against its list price per second of video, on a logarithmic axis. Prices are recorded manually from the providers’ published rates. Error bars are 95% confidence intervals of the score. Models ringed in teal and joined by the step line form the Pareto frontier: no other model has both a lower cost and a higher score. The remaining models are faded. 6 models without a cost in USD per video second are not shown.
We use crowd intelligence to deliver real, diverse human feedback at scale for model evaluation and post-training.