3D/4D Consistency

The consistency_3d_4d dimension asks whether a generated rollout admits a stable spatiotemporal geometric explanation in three or four dimensions. It aggregates official metric scores into per_dimension_summary.json as one of WorldAtlas’s top-level evaluation axes.

This is distinct from perceptual quality, instruction following, or physical-dynamics scoring. A prediction can look sharp and follow the requested prompt while still warping backgrounds, breaking depth layering, or failing to close as one continuous viewpoint. 3D/4D consistency therefore measures geometric self-consistency—whether camera motion, depth structure, cross-frame correspondence, and short-horizon motion fields can be explained jointly under one geometric hypothesis.

The dimension spans four geometry families. Multiview representation is covered by met3r_consistency, which scores whether protocol anchor views describe the same scene without requiring pose ground truth, and by the reconstruction pair geometric_consistency and photometric_consistency, which test whether the prediction can be recovered as a depth-and-pose point cloud and reprojected across views. Classical epipolar geometry uses epipolar_error, epipolar_inlier_rate, and feature_match_coverage to ask whether adjacent frames admit a stable fundamental-matrix / Sampson explanation. Camera and depth structure includes reprojection_error for dense SLAM-reprojected covisible points. Motion-field consistency is read by optical_flow_aepe through forward–backward optical-flow round trips.

WorldAtlas applies strict eligibility rules throughout this dimension. On the current public surface, met3r_consistency, reprojection_error, optical_flow_aepe, geometric_consistency, and photometric_consistency are official when they succeed; the epipolar trio is also official. Legacy aliases identify their documented metric implementation; suite selection and aggregate weights determine which columns are combined. Reconstruction failure, unreadable media, or unmet pairing constraints yield normalized = null rather than a forced zero; an empty MEt3R overlap mask is an explicit worst-case official score. Read 3D/4D consistency together with action-control, physics, and quality metrics: a unified geometric explanation does not imply that the model obeyed its controls or that object motion is physically plausible.