This benchmark ranks world and video generation models on whether they can move the camera to a place defined by the scene itself, such as behind the sofa, into the driver’s seat or to the dog’s point of view, starting from a single still image. We wrote 220 camera instructions for 50 scenes, each placing the camera relative to what its scene contains rather than along a stated direction, held fixed to ensure reproducibility of the results. The rankings are updated continuously from 955,708 responses across 158,383 pairwise human-preference matches. All data and judgments are public; see the methodology for details on evaluation and scoring.
We use crowd intelligence to deliver real, diverse human feedback at scale for model evaluation and post-training.