The usefulness of world and video models in robotics, autonomous driving and games depends heavily on how well they model physics, which has made physical accuracy a growing focus. Physics-IQ (Motamed et al., 2026) and its refined version Physics-IQ Verified (Rädsch et al., 2026) measure this automatically: a model continues a real recording of a physical experiment, and its video is compared with what actually happened using hand-crafted metrics on motion and pixels. We build on this work independently: on the same 198 real recordings across 66 physical scenarios, we rank models by human judgement instead of automated metrics. We evaluate 26 video and world models, which are each given an input frame taken from a real recording and a prompt describing the scene, without revealing the expected outcome.The Blind mode presents two generated videos and asks human raters which of them looks more realistic.The Referenced mode additionally shows the original recording and asks which video reproduces it more faithfully.Using 790,940 human judgements and 132,890 pairwise comparisons, we compared 26 models. In our overall leaderboard, which weights both modes equally, FLUX 3 Max (post-trained by fal on Black Forest Labs' FLUX 3) emerged as the winner, followed by Gemini Omni 1.1 Flash and Dreamina Seedance 2.5.
We use crowd intelligence to deliver real, diverse human feedback at scale for model evaluation and post-training.