Any Time, Any Space: A World-State-Centric Framework for Agentic Video Generation

• Nonconsecutive return shots must preserve established scene layout, while event replays must reproduce the earlier actions and temporal progression from a different viewpoint.Source: §1 · p. 1
• ATAS grounds screenplay requests in persistent scene state with shared spatial structure and event state that retains dynamic observations with temporal context.Source: §1 · p. 2
• Generated clips remain candidates until reconstructed observations are validated, and film suitability is assessed separately from whether evidence updates persistent state.Source: §1 · p. 2
• Across four generators, ATAS improves every reported cross-shot consistency measure and achieves the highest aggregate score in each backbone block.Source: §1 · p. 3
• ATAS outperforms evaluated agent and 3D-guided methods on all three spatial-revisit metrics, leads three of five event-recapture criteria, and has the highest mean human rating on all six criteria.Source: §1 · p. 3

Video Generation

• Text-to-video, long-video, and camera- or geometry-guided methods improve local rendering or control, but do not specify how observations are registered, retained, or queried in later shots.The passage describes text-to-video and image-to-video models [2] [3] [4] [5] [6] [7] [8] [9] [10] [11], long-video methods [12] [13] [14], and camera-, sparse-control-, and geometry-guided methods [17] [18] [19] [20] [21].Source: §2 · p. 3

Agentic Video Generation

• Agentic systems add script parsing, shot planning, reference selection, and review [22] [23] [24] [25] [26] [27] [28] [29] [30] [31], yet their editorial context typically lacks shared coordinates, event-time indexing, and an admission rule.Active view selection supports spatial reasoning [32]; ATAS separates compilation, retrieval, generation, and state update in a plan-act-observe loop [33].Source: §2 · p. 3

Explicit Memory for Video Generation

• Explicit-memory and 3D methods provide geometry, dynamic view synthesis, or world assets [34] [35] [36] [37] [38] [39] [40] [41] [42] [43] [44] [45] [46] [47] [48] [49] [50] [51] [52] [53] [54] [55], but the passage distinguishes ATAS by its shot-local evidence, event-time addressing, and state-update decision interface.ATAS separates static scene memory from event-local dynamic evidence and evaluates observations before persistent-state updates.Source: §2 · p. 3

3 Method

• ATAS keeps spatial and temporal evidence in a persistent world state that is queried before each shot and updated with observations that pass validation.The state comprises scene and event memories.Source: §3 · p. 3, §3.1 · p. 3

3.1 Terms and Workflow

• Each request specifies a description, scene, spatial anchor, camera, time interval, and optional event, with scene and event memories supplying relevant references.The workflow conditions the generator on both the shot prompt and context-dependent prompt.Source: Figure 2 · p. 4

3.2 World-state Compilation

• The compiler recursively splits a screenplay by scene, spatial and temporal relations, event references, and camera requirements until each request fits the short-video budget.Requests are processed in order, and a validated next state conditions the following request; rejected observations remain film but not state evidence.Source: §3.2 · p. 4

3.3 Two-Stage Memory: Scene and Event Memory

• Scene memory stores accepted static geometry registered to the source observation’s coordinate frame, while event memory stores dynamic evidence with scene, event, and time information.An event query retrieves compatible records; no compatible record yields an empty result, and dynamic evidence remains separate from static geometry.Source: §3.3 · pp. 4–5

3.4 Shot Generation from Queried State

• The shot prompt supplies local intent and place and time, while the context-dependent prompt supplies scene references and, when needed, event references to the video generator.The generation interface is kept fixed across comparisons.Source: §3.4 · p. 5

3.5 Observation Registration and State Update

• Film review, registration, and state-update checks separately determine delivery, alignment to the scene frame, and admission of an observation to persistent state.A clip accepted for film can remain usable even when registration or state admission fails.Source: §3.5 · pp. 5–6

3.6 Two Persistence Queries

• Spatial revisit queries the earlier scene and return anchor, whereas historical event recapture retrieves records for an earlier event and time interval after intervening shots.The two queries test recovery of static spatial evidence and addressability of a time-local event, respectively.Source: §3.6 · p. 6

4.1 Experimental Setup

• Evaluation uses official VBench, ST-Bench, and ViStory inputs alongside task-specific revisit and recapture evidence; human raters score anonymized whole-video outputs on a five-point scale.Source: §4.1 · pp. 6–8

4.2 Main Results

• ATAS leads event identity (0.750), temporal order (0.688), and viewpoint realization (0.650), ties ReCA on character and object consistency (0.917), and trails on action-state preservation (0.688 versus 0.700).Source: Table 3 · p. 8, §4.2 · p. 9
• ATAS reaches 15.968 PSNRC, 0.385 SSIMC, and 0.284 LPIPSC, outperforming the compared multi-shot and 3D-guided methods on spatial revisit.Source: Table 2 · p. 8, §4.2 · p. 8
• ATAS leads every reported aggregate on MiniMax-H3, Wan 3.0, and HappyHorse 1.0; on Wan 2.7 it leads ST-Bench and both ViStory consistency measures but trails VBench.Source: Table 1 · p. 7, §4.2 · p. 8
• Human raters give ATAS the highest mean score on all six criteria; its highest listed ratings are script faithfulness (4.40), character consistency (4.22), and visual appeal (4.10).Source: Figure 5 · p. 9, §4.2 · p. 9

4.3 Ablation Study and Experimental Analysis

• Adding memory raises ViStory average from 0.478 to 0.612; recursive compilation then raises event score to 0.711 and C-Top10 to 0.684.Source: Table 4 · p. 9, §4.3 · p. 10

4.2 Main Results

• Retries fall from 1.00 to 0.54 per shot and retry-adjusted delivery time from 359 to 339 seconds, while per-story runtime rises from 51.9 to 64.9 minutes.Source: §4.2 · p. 9, Figure 6 · p. 10

4.3 Ablation Study and Experimental Analysis

• Across the memory ablation, C-Top10 rises from 0.536 to 0.719, event score from 0.452 to 0.739, and LPIPSC falls from 0.341 to 0.284.Source: Table 5 · p. 10, §4.3 · p. 10

5 Conclusion

• ATAS compiles screenplays into shot requests grounded in explicit scene, camera, and event-time references, then updates persistent state only from validated observations.Source: §5 · p. 10
• Ablations show complementary contributions from scene memory, event memory, and recursive compilation, while efficiency analysis finds fewer retries at added compute cost.Source: §5 · p. 11
• By keeping locations and prior events addressable throughout a story, ATAS provides a foundation for interactive filmmaking, episodic content creation, and previsualization.Source: §5 · p. 11

Contribution and Acknowledgements

• The contribution statement credits the authors’ roles in project leadership, implementation, validation, screenplay design, manuscript writing, experimental design, and feedback, and acknowledges participants and tool developers.Source: Contribution and Acknowledgements · p. 11

Appendix Contents

• The appendix covers screenplay compilation, scene and event memory, benchmark evaluation, controlled analyses, human judgments, and extended story rollouts.Source: Appendix Contents · pp. 17–18

A Evaluation Protocols and Benchmark Details

• Benchmark breakdowns use the same backbone and method grouping as Tab. 1, separating local rendering quality from cross-shot behavior rather than forming a new training objective.Source: A Evaluation Protocols and Benchmark… · p. 18

A.1 Source-grounded Diagnostic Evaluator

• The source-grounded evaluator scores five distinct event-recapture axes from source frames and a camera request, but cannot establish pixel-level agreement with an unobserved target video.Source: §A.1 · p. 18

A.2 VBench Dimensions

• VBench breakdowns show different local strengths across backbones: Wan 3.0 leads subject and background consistency and dynamic degree, Wan 2.7 flicker and motion smoothness, and MiniMax-H3 aesthetics.Source: §A.1 · p. 19, §A.2 · p. 19

A.3 ST-Bench Metrics

• ATAS’s strongest ST-Bench measures are cross-shot consistency, especially C-Top10; Wan 3.0 reaches 0.6844 on C-All and 0.7187 on C-Top10.Source: §A.3 · p. 19
• Wan 3.0 reaches 0.2415 on PF-G and 0.2362 on PF-S, smaller and more uniform prompt-following changes than its consistency changes.Source: §A.3 · p. 19
• C-Top10 exceeds C-All by 0.0069–0.0379 across ATAS backbones, while adjacent ViStory breakdowns distinguish within-story from cross-shot consistency dimensions.Source: §A.3 · p. 19, Table C · p. 20, Table D · p. 20

A.4 ViStory Metrics

• Wan 3.0 leads identity, appearance, and scene consistency, and every ATAS backbone exceeds the reference aggregate in Tab. D.Across Tab. C, Wan 3.0 also has the strongest appearance and temporal self-consistency and the highest aggregate; the other ATAS backbones remain above the reference aggregate.Source: §A.4 · p. 20

B World-state compiler and bounded shot requests

• The compiler recursively resolves scene transitions, time relations, historical events, and camera requirements into bounded requests before rendering.Each request carries a local description, scene coordinate, target camera, time interval, and optional event information; later requests wait for validated observations before using updated state.Source: B World-state compiler and bounded… · pp. 20–21

C Memory Construction and Query

• Scene memory stores registered static evidence, while event memory retains dynamic observations tied to a source shot, event, and time interval.The two memories are queried separately and combined for historical-event requests; reconstruction and registration produce candidate observations after generation, while retrieval supplies references before the next generation.Source: C Memory Construction and Query · pp. 21–22
• From rollout start to end, scene size grows from 33.54 to 49.60 million units and event records from 1.0 to 4.8, while query and update time rises from 93.0 to 103.5.Storage rises from 0.95 to 2.44 in the reported units; these checkpoints describe the reported configuration, not a general scaling law.Source: C Memory Construction and Query · p. 22

Memory capacity and update sensitivity

• Reducing scene capacity lowers spatial quality, while shortening event history reduces event quality: C-Top10 falls to 0.638 and LPIPSC is 0.301, while the event score falls to 0.608.Relative to full update, sparse updates reduce C-Top10 from 0.719 to 0.685 and event score from 0.739 to 0.732; shortening event history leaves LPIPSC almost unchanged.Source: D Sensitivity and Mechanism Analysis · pp. 22–23
• Update frequency and scene capacity are distinct controls: more frequent updates cannot fully compensate for a smaller scene store, and a large scene store cannot replace missing event history.Shortening event history affects event quality more than spatial quality, consistent with the separation between static scene evidence and time-local event evidence.Source: D Sensitivity and Mechanism Analysis · pp. 22–23

Memory freshness and staleness

• C-Top10 declines from 0.736 with zero intervening shots to 0.651 after five or more, while LPIPSC rises from 0.276 to 0.306 and event score falls from 0.762 to 0.678.The bins do not identify whether the loss comes from scene changes, event density, or accumulated registration error.Source: D Sensitivity and Mechanism Analysis · p. 23

State update check

• The default state-update check accepts 67.4% of candidate updates and yields C-Top10 0.719 and event score 0.739, outperforming both permissive and strict settings on those measures.Permissive admission accepts 90.0% but yields 0.696 and 0.708; strict admission accepts 36.0% and yields 0.668 and 0.682. The default is not established as universally optimal.Source: D Sensitivity and Mechanism Analysis · p. 23

Human ratings

• ATAS has the highest mean rating in all six criteria, ranging from 3.62 to 4.40, while ReCA, the strongest baseline, ranges from 2.62 to 3.67.Ratings of at least four points range from 57% to 95% for ATAS and 14% to 64% for ReCA; agreement coefficients range from 0.58 to 0.72.Source: E Human Evaluation · pp. 23–25
• Narrative coherence and physical-law judgments remain harder than script faithfulness and character consistency.Raters watch anonymized clips in randomized order and score the whole video on a fixed five-point scale.Source: E Human Evaluation · pp. 23, 25

Automatic diagnostics and human judgments

• Automatic-diagnostic correlations with human ratings range from 0.41 for temporal phase to 0.61 for event identity; human–human ceilings are higher at 0.67–0.80.The passage supports treating automatic metrics as structured diagnostics rather than complete substitutes for human judgment.Source: E Human Evaluation · p. 25

Qualitative baseline audit

• In record-shop and bakery sequences, methods differ in whether room layout, object placement, and character state remain consistent after intervening shots.The matched-frame audit compares shot continuation, scene layout, object state, and character consistency; a clip may remain locally plausible while drifting in layout or entity state.Source: F Qualitative Results · p. 25, Figure I · p. 26

Representative rollout timeline

• The rollout timeline tracks scene changes, a spatial revisit, and an event-recapture request alongside memory conditions and returned observations.It shows when a later request queries prior state and when a returned observation is written back.Source: F Qualitative Results · p. 27

Representative multi-story results

• The multi-story grid presents timestamped frames from six stories spanning different settings, people, objects, and camera patterns.The figure illustrates the range of scenes, actions, and camera requests used in the qualitative evaluation.Source: F Qualitative Results · p. 27

G Long-Horizon Result Gallery

• Across three full-length films, later views retain established layouts, objects, and changes, demonstrating generated continuity rather than physical simulation of unobserved intervals.Source: G Long-Horizon Result Gallery · p. 27

G.1 Saturday Laundry in the Flat

• On returns, living-room furniture and the mug remain in place, while the laundry-room setup persists and the started washer remains shut until collection.Source: §G.1 · p. 30

G.2 Winter Fire in the Cabin

• The kitchen return retains the heart in the frosted window, and the living-room return shows the fire she lit still burning and filling the hearth.Source: §G.2 · p. 33

G.3 Running a Bath on a Winter Evening

• On return, the closed laptop remains on the bed; the partly filled tub is full of pink water, and the candle remains burning.Source: §G.3 · pp. 36–37