Spatial revisit
Compare a generated return view with a target image under a common-visibility mask. ATAS reports the best value on PSNRC, SSIMC and LPIPSC among the listed methods.
Results / 02
Measure the story
between the shots.
ATAS evaluates whether a generated world remains addressable after the film has moved on—not only whether one isolated clip looks good.
Headline result
Consistency is where the gap opens.
Across four rendering backbones, the paper reports the highest aggregate benchmark score for ATAS in each block. The clearest gains appear on measures that cross shot boundaries.
Wan 3.0 average0.734ATAS aggregate score
Cross-shot consistency0.584ViStory-Cross on Wan 3.0
Spatial revisit15.968PSNRC reported for ATAS
Human evaluation6 / 6highest mean across all criteria
Numbers are reproduced as reported in the paper and project page. They are not a promise of future performance or a substitute for reproducing the evaluation.
Standard benchmarks
Four views of quality.
The evaluation combines local visual quality with measures that ask whether identity, appearance, scene layout and temporal state survive across a story.
| Benchmark | Question | What to notice |
|---|---|---|
| VBench | Does the clip render well? | Local appearance, subject consistency and motion. |
| ST-Bench | Do shots stay consistent? | All-pair and top-ten cross-shot consistency. |
| ViStory-Self | Does one story hold together? | Identity, appearance and temporal self-consistency. |
| ViStory-Cross | Do story elements survive across shots? | Cross-shot identity, appearance and scene consistency. |
Targeted protocols
Return to a place.
Replay an event.
Two targeted tests make the idea concrete: a later camera returns to an earlier space, or a past event is requested again after intervening shots.
Event recapture
Use source-grounded evidence to judge event identity, action and state, temporal order, character and object consistency, and viewpoint realization.
Human evaluation
The whole film still matters.
Raters watch anonymized long videos and score six criteria from 1 to 5: visual appeal, script faithfulness, character consistency, background consistency, physical law and narrative coherence.
The paper also reports that automatic diagnostics correlate with human ratings, but remain below the human–human agreement ceiling.
Limitations
A strong result with visible open questions.
- Memory freshness: reported quality gradually declines as more shots intervene after the latest accepted update.
- Admission trade-off: accepting too much noisy evidence or rejecting too much useful evidence can both reduce downstream quality.
- Exact action-state reproduction: event recapture remains harder than identifying the event or viewpoint.
- Evaluation boundary: automatic metrics are structured diagnostics; they do not replace human judgment.