Long Sequence / Memory

The long_sequence dimension asks whether a generated rollout keeps what it has already shown, and whether a place is still itself after the camera leaves and returns. It aggregates official metric scores into per_dimension_summary.json as one of WorldAtlas's top-level evaluation axes.

Dimension key

long_sequence

Rolls up into per_dimension_summary.json

Families

3

Persistence, long-horizon quality, closed-loop memory

Core question

State persistence

Do later observations stay consistent with earlier ones over a long horizon?

Scope

Metrics grouped by persistence, long-horizon quality, and closed-loop memory.

Long-horizon quality

No-reference aesthetic and imaging scores over the rollout. These do not require a closed camera path.

• long_sequence_aesthetic_quality
• long_sequence_imaging_quality

Persistence

Already-seen regions and objects should survive as the rollout continues. The Long-Sequence backend also supplies motion smoothness and dynamic degree. These do not require a closed camera path.

• long_sequence_motion_smoothness
• long_sequence_dynamic_degree

Closed-loop memory

The memory_loop suite: one still image, a 60-second rollout, and a geometrically closed camera path. Revisit references are earlier frames the model itself generated at the same estimated viewpoint.

• memory_return_gate — official
• memory_scene_f1
• memory_revisit_fidelity
• memory_geometric_closure
• memory_rendering_recovery
• memory_half_life — official

Alternative measurement protocols

Split on ground-truth action indices or first and last protocol frames. These are official documented metrics; compare only results from the same protocol.

• scene_memory_f1
• memory_revisit_consistency

Reading notes

• This is distinct from action control, geometric consistency, or perceptual quality. A prediction can follow the requested path and still forget the room it left; it can also look sharp frame-to-frame while walls and furniture change once the viewpoint comes back.
• memory_return_gate decides whether the camera actually left and returned. Official closed-loop scores come from memory_scene_f1 for point-cloud overlap, memory_geometric_closure for pose/feature consistency, and memory_revisit_fidelity and memory_rendering_recovery for recovered appearance. Pairs are read from estimated poses, not from symmetric time indices.
• The leaderboard reports atomic metrics without a mixed-protocol memory overall. memory_scene_f1 is a task mean: valid-input return-gate failures count as zero; genuinely unavailable measurements remain missing and coverage is displayed. Other memory columns retain their own conditional-measurement semantics and provenance; a scene-only update does not revalidate them.
• memory_half_life is an exploratory fidelity-decay fit: a reliable numeric fit is an official leaderboard result. It is not a validated measure of causal memory lifetime. The official gate metric is reported as a pass rate.
• Read long-sequence scores together with action control: a model that never walks back has no revisit, not a forgotten scene.