Quality

The quality dimension asks whether a generated rollout is visually stable, clear, and natural, and whether it remains reconstructively similar when a paired reference exists. It aggregates official metric scores into per_dimension_summary.json as one of WorldAtlas’s top-level evaluation axes.

This is distinct from instruction following, geometric self-consistency, or physical-dynamics scoring. A prediction can look sharp and well-styled while still violating the requested scene, breaking 3D structure, or drifting physically. Quality therefore measures visual fidelity—style, perceptual naturalness, low-level appearance stability, and, when references exist, reconstruction-style agreement.

The dimension spans two quality families. Style and perceptual quality is covered by style_consistency, perceptual_quality, image_quality, and hpsv3_norm, plus the suite-level distribution metric jedi. Appearance stability uses brightness_distribution_consistency, color_temperature_constraint, and sharpness_retention for low-level photometric drift, together with temporal_flickering, background_consistency, and segment_continuity for flicker, scene-appearance drift, and unintended cuts. rollout_aesthetic_drift and rollout_imaging_drift are official rollout-length quality metrics. Long-horizon no-reference scores long_sequence_aesthetic_quality and long_sequence_imaging_quality are documented under Long Sequence / Memory.

WorldAtlas applies strict eligibility rules throughout this dimension. The seven complete official image_static quality metrics—sharpness_retention, perceptual_quality, style_consistency, brightness_distribution_consistency, color_temperature_constraint, image_quality, and hpsv3_norm—enter per_metric_summary and the equal-weight quality total. Long-horizon drift metrics are official within their configured suites. Legacy aliases are not additional copies of the canonical metrics. These low-level slots are inspired by iWorld-Bench quality axes, not reproductions of unpublished iWorld-Bench code. Read quality scores together with action-control, consistency, and physics metrics: strong appearance alone does not imply that a world model followed its controls, remembered the scene, or obeyed physical dynamics.