WorldWeave: Growing Persistent Geometric Worlds for Video Generation

• World models based primarily on images, video, or finite observation histories still face challenges maintaining explicit, persistent structural memory across expansion and revisitation.Source: §1 · p. 2
• WorldWeave stores terrain, regional organization, asset identities, transforms, and boundary interfaces in a shared metric 3D world, committing only validated additions.Each expansion builds an independent candidate from committed state, leaving existing structural records unchanged.Source: §1 · p. 2
• Diffusion-based terrain outpainting, agent-guided hierarchical planning, deterministic geometry compilation, and bounded local revision construct connected scene additions from user intent and terrain evidence.Accepted candidates are committed as new world versions, and planned camera trajectories query the state for depth-conditioned video synthesis without RGB write-back.Source: §1 · p. 2
• World geometry and relations remain in the shared world while appearance is synthesized per viewing request, supporting structural consistency across trajectories and revisits.The paper evaluates structural consistency and revisit memory alongside camera compliance and visual quality.Source: §1 · p. 2

• Video world models use approaches including geometry conditioning, causal caching, frame-state retrieval, geometric caches, and camera-indexed world-state updates; WorldWeave maintains memory through validated scene construction.The related work contrasts its committed-geometry reads with systems that update state from generated frames.Source: §2 · pp. 2–3
• Expandable-world methods grow scenes through layout grids, view synthesis, geometric integration, semantic relations, shared scaffolds, or spatial blocks; WorldWeave preserves committed structure and inherits terrain and scene interfaces.Source: §2 · p. 3
• Agent-guided construction includes procedural environments, language-driven programs, guided layouts, region-aware terrain, and feedback-based refinement; WorldWeave adds inherited interfaces, deterministic compilation, and bounded revision before commit.Source: §2 · p. 3
• Terrain approaches include hydrology-based modeling and learned sketch-, texture-, or text-conditioned generation; WorldWeave generates metric chunks against committed neighbors and joins them with new-side height and derivative constraints.Source: §2 · p. 3

3.1 Persistent Structural Memory

• Committed versions store elevation and boundary derivatives, regional organization and object relations, asset identities and transforms, and interfaces specifying terrain attachment and road or water connections.Expansions form independent candidates from the committed version; camera queries select an existing version without changing its structural records.Source: §3.1 · p. 3

3.2 Continual Metric Terrain Expansion

• Each 512 × 512 terrain chunk is generated against one to three committed neighbors using reference-relative residual channels, then joined with new-side height and derivative constraints.New connector cells inherit boundary heights and derivatives; corrections vanish at their inner support boundaries while committed geometry remains fixed.Source: §3.2 · pp. 4–5

3.3 Hierarchical Scene Planning and Compilation

• Terrain evidence and inherited interfaces guide region roles, district extents and access, and asset relations; deterministic compilation realizes plans as metric layouts and infrastructure connections.Catalog attributes and geometric predicates link semantic choices to spatial constraints and identify failed relations for revision.Source: §3.3 · p. 5

3.4 Bounded Revision and Validated Commit

• Program checks and matched-view geometry previews guide targeted local revisions; review evidence is regenerated after edits, and passing candidates are atomically published as new versions preserving committed records.Candidates that exhaust the revision budget are rejected.Source: §3.4 · pp. 5–6

3.5 Read-only Queries for Video Generation

• Camera requests are checked for clearance and viewpoints, then geometry queries return calibrated metric depth; depth guides spatial structure while optional style references control appearance, with no RGB write-back.The queried surfaces retain shared world coordinates across views, supporting geometric correspondence through viewpoint changes and revisits.Source: §3.5 · p. 6

4.1 Experimental Setup

• The benchmark covers 120 tasks spanning forward exploration, viewpoint changes, and occlusion or off-screen return, with videos evaluated for 15 seconds at 12 fps.Baselines include six open-source world models and four video baselines; all receive the same task and scene-description prompt, while WorldWeave additionally receives queried depth.Source: §4.1 · p. 6

4.3 Structural Consistency and Camera Compliance

• WorldWeave achieves the lowest MC-GeCo and MC-MEt3R errors, 0.0573 and 0.122, improving over MiniMax-H3 by 9.6% and 12.5%, respectively.WorldWeave also obtains the highest SMC score, 2.30 versus 2.21 for runner-up Echo-WM, and ranks first in Cam at 72.3, 2.49 points ahead of Seedance 2.0.Source: §4.3 · p. 7

4.2 Main Quantitative Results

• WorldWeave ranks fifth in IQ and second in AQ, while leading MN-MS, MC-GeCo, SMC, Cam, MC-MEt3R and MC-GeoCon.Table 1 compares visual quality, structural consistency, and camera control; WorldWeave improves all eight measures over MiniMax-H3, including SMC (+426.7%) and Cam (+42.2%).Source: §4.2 · p. 7, Table 1 · p. 8

4.3 Structural Consistency and Camera Compliance

• On 120 matched tasks, WorldWeave’s mean optical-flow amplitude exceeds six comparison methods by 8.1–24.1% while it maintains leading structural-consistency scores.The comparisons are SANA-WM, EVOKE-Turbo, AlayaWorld, Seedance 2.0, Wan3.0, and Kling 3.0.Source: §4.3 · p. 8

4.4 Ablation Studies

• Multiscale residual channels reduce VAE round-trip RMSE to 0.158 m, and the full elevation codec reaches 0.136 m; canvas conditioning reduces boundary-height gaps by 65.4–76.1%.The remaining canvas-conditioned gaps are 0.264–0.368 m for nominally 256 m chunks; slope continuity improves in two- and three-neighbor settings, while single-neighbor input favors multi-image conditioning.Source: §4.4 · p. 8, Table 2 · p. 9
• Full derivative inheritance reduces both terrain-joining residuals below numerical tolerance, while independent slope taper reduces mean whole-target modification from 0.2088 to 0.0825 m (60.5%).All joining variants connect every target and preserve valid existing heights and derivatives.Source: §4.4 · p. 8
• Changing the planning model alters settlement density and infrastructure, while both planners continue inherited roads and retain shared terrain and cross-region organization.Across five scene–trajectory pairs and five seeds, seed-wise mean IQ, AQ, and MN-MS have coefficients of variation below 0.4%; the three geometric metrics range from 3.1% to 6.3%, while event-based SMC and Cam vary more strongly.Source: §4.4 · p. 9

4.5 User Study

• The user-study table reports ratings on a 1–5 scale, with higher scores indicating better ratings.The supplied table caption identifies the rating scale but does not provide rating outcomes.Source: Table 3 · p. 9

5 Conclusion

• On 20 selected cases, WorldWeave leads Echo-WM in camera, object, and structure/texture consistency ratings.WorldWeave scores 4.57, 4.73, and 4.52, versus Echo-WM’s 3.92, 3.42, and 3.32.Source: §5 · p. 9
• WorldWeave’s read-only camera queries and incremental construction extend an explicit world while preserving committed structure; evaluated methods show leading structural-consistency and camera-compliance measures.The conclusion also reports competitive visual quality and higher realized motion than most baselines.Source: §5 · p. 9
• The current system focuses on static geometry, and depth guides video synthesis rather than imposing a hard rendering constraint.Future work targets dynamic interactions, long-horizon appearance memory, and adaptive spatial detail for larger worlds.Source: §5 · p. 9

AI Use Statement

• Generative AI assisted with drafting, writing organization and polishing, and literature retrieval; the authors remain responsible for the final manuscript’s accuracy and integrity.Source: AI Use Statement · p. 9

Ethics Statement

• The terrain and geometry assets come from authored virtual environments, not surveyed real-world geography; generated worlds should not be treated as verified real locations.The statement also calls for disclosure of synthetic content and preservation of provenance.Source: Ethics Statement · p. 10

Reproducibility Statement

• The reproducibility statement locates construction and observation interfaces in Section 3 and principal evaluation comparisons in Section 4, with implementation and protocol details across Appendices A–F.Source: Reproducibility Statement · p. 10

A.1 Ablation Protocol

• The planning-model substitution fixes terrain, intent, assets, compiler, revision budget, and cameras while comparing relation realization, geometric validity, and revision cost.The default GPT-5.6-Sol planner is compared with Qwen3.8-Max.Source: §A.1 · p. 14

A.2 Visual Comparison of Terrain Conditioning

• With matched terrain examples, spatial-canvas conditioning maintains more coherent relief across marked interfaces than separate-image conditioning.The comparison complements the quantitative boundary-continuity results.Source: §A.2 · p. 14

A.3 Planning-model Comparison

• With identical terrain and the existing region unchanged, the alternative planner favors wooded open space and dispersed settlements, while the default produces denser settlement and roads.Both retain the inherited terrain interface and organized connections.Source: §A.3 · p. 14

A.4.1 Controlled Comparison

• The controlled comparison fixes five scene–trajectory pairs, world version, first frame, prompt, depth, and non-seed settings while varying seeds 0, 1, 2, 42, and 100.Table 4 retains all 25 controlled outputs alongside five main-experiment references.Source: §A.4.1 · p. 14

A.4.2 Metric Variation and Visual Outcomes

• Aggregate IQ, AQ, and MN-MS have CVs of 0.39%, 0.40%, and 0.05%; geometric-score CVs range from 3.11% to 6.33%, though averages do not imply every video is equally stable.Table 5 reports seed-wise means over five tasks and across-seed dispersion, excluding the five main-experiment references; SD is population standard deviation and CV is SD divided by the mean.Source: §A.4.2 · p. 14, Table 5 · p. 16
• One of 25 controlled outputs, the mountain scene with seed 100, has black-sky corruption; this small study does not establish how often it occurs in deployment.The affected video remains in the statistics, and its outcome is consistent with—but does not isolate—dark depth-conditioning regions transferring into sky appearance.Source: §A.4.2 · p. 14, Table 4 · p. 16
• SMC and Cam vary more than computational measures, with CVs of 27.61% and 15.60%; event decisions and evaluator sensitivity contribute to their scores.The experiment varies generation seeds, not repeated judgments of the same video, so it does not attribute all dispersion to evaluator randomness.Source: §A.4.2 · p. 14
• Figure 6 shows terrain completions using a spatial canvas versus separate images, with shared known terrain and dashed context–target interfaces.Gray regions indicate unavailable terrain.Source: Figure 6 · p. 15
• Figure 7 places the shared existing region on the left and new region on the right for alternative and default planner layouts, using identical terrain and drawing conventions.Source: Figure 7 · p. 16

B.1 Metric Definitions

• SMC evaluates identity, structure, neighbor relations, and texture when a designated target disappears and returns; unobservable texture grades are omitted.A method-blind evaluator grades sampled return views against an earliest usable reference, with caps for low identity, structure, or neighbor grades.Source: §B.1 · p. 17
• Cam evaluates whether requested translation, turning, framing, absence, and return clauses are executed, using camera poses or visual evidence as appropriate.Event clauses receive binary scores and sustained clauses the fraction of intervals satisfying them.Source: §B.1 · p. 17
• GeCo combines motion and depth residuals, MEt3R compares reprojected image features, and GeoCon-Bench measures correspondence inliers and geometric fitting error.Source: §B.1 · p. 17

B.2 Motion-Calibrated Consistency and Smoothness

• Raw disagreement can decrease when a model barely changes its view, so the evaluation normalizes geometric error by realized motion and measures camera compliance separately.The relative residual uses co-visible motion and depth evidence; newly exposed content is excluded from correspondence-failure scoring.Source: §B.2 · pp. 17–18
• Task-relative normalization divides each video’s error by realized motion relative to a fixed task-wise reference; the same calibration is applied to VBench smoothness error.Scores are aggregated from video-wise ratios, and higher values are better when interpreted jointly with geometric consistency and camera compliance.Source: §B.2 · pp. 17–18

C User-study Protocol

• The user study uses 20 matched scene–trajectory cases, seven methods per case, and 1,308 criterion scores from 436 video evaluations.All 140 videos received at least three ratings; cases were selected using automatic metrics and visual screening.Source: C User-study Protocol · p. 18
• WorldWeave has the unique highest object-consistency mean in 19 of 20 cases, and the highest or tied-highest camera and structure/texture means in 16 and 18 cases.Ratings use a 1–5 scale and average cases equally.Source: C User-study Protocol · p. 19

D.1 Context Reference and Coordinate Convention

• A least-squares context plane is refined by reweighted fits; with one neighbor, its normal trend tapers away from the interface to prevent indefinite linear extrapolation.The context bands are 64 samples wide, and target elevations are excluded from the fit.Source: §D.1 · p. 19

D.2 Redundant Encoding and Robust Metric Decoding

• Elevation residuals are redundantly encoded across RGB channels and spatial repetitions, then robustly decoded to the metric grid despite channel disagreement.Saturated channels receive less weight during inversion, followed by a fixed robust projection.Source: §D.2 · pp. 19–20

D.3 Conditioning, Supervision and Boundary Acceptance

• Terrain conditioning separates known context, target, and unavailable locations, while supervision covers valid target samples and joining checks shared interfaces.Boundary corrections apply on the new side, preserving inherited old surfaces and reconciling corners on new-owned nodes.Source: §D.3 · p. 20

D.4 Ownership-aware Continuous Joining

• Bicubic Hermite connector cells inherit height and derivatives along preserved edges, assigning gaps between old sample centers and uncovered corners to the new connector.Height and slope corrections are applied on the new side, with normalized corner weights.Source: §D.4 · p. 20
• Slope taper reduces mean modification by 0.1263 m relative to fixed C1 joining, while the largest pointwise change remains 35.37 m.Exact boundary closure follows from derivative inheritance and does not imply uniformly small modifications or gentle connector slopes; the experiment retains all 256 targets.Source: §D.4 · p. 20

E.1 Terrain Evidence and Spatial Responsibilities

• Terrain evidence, inherited interfaces, and asset capabilities inform regional roles, while region and district layers establish spatial responsibilities before placement.Roads, water corridors, vegetation overlays, and boundary ports carry separate typed constraints.Source: §E.1 · p. 21

E.2 Typed Relations and Asset Transforms

• Typed relation graphs encode asset roles and connections, and deterministic transforms align connector endpoints while rejecting forbidden span changes.Other asset families follow their own support and orientation contracts.Source: §E.2 · p. 21

E.3 Geometric Validation and Repair Attribution

• Geometric validation checks asset-specific clearance and road access, and failed predicates retain instance, relation, and district identifiers to localize revisions.Support relations allow intended contact, while clearance applies to incompatible object pairs.Source: §E.3 · pp. 21–22

F.1 Bounded Candidate Revision

• Candidate publication requires all geometric predicates and both region and compiled-scene reviews to pass; bounded edits preserve unaffected decisions and reject failed candidates when revision budget runs out.A rejected candidate leaves the previous version available.Source: §F.1 · p. 22

F.2 Versioned Publication

• Publication binds validated terrain, plans, transforms, and interfaces into a version; existing instance identities and transforms persist across additions.A query retains its selected version throughout a trajectory.Source: §F.2 · p. 22

F.3 Camera Geometry and Video Conditions

• Camera queries use the selected world’s terrain and transformed assets to determine visibility, hit identity, and depth; RGB synthesis operates without a write-back path.Pixels without valid hits carry an explicit validity mask, and the adapter preserves depth convention and calibration.Source: §F.3 · p. 22

F.4 Relation to Joint-Embedding Predictive Architectures

• WorldWeave’s connection to JEPA is conceptual: it constructs and preserves a geometrically addressable world rather than adopting JEPA’s predictive-representation training objective.Its persistent state includes metric terrain, asset instances, semantic organization, and cross-region interfaces.Source: §F.4 · p. 23

G Additional Visual Comparisons

• The farmstead sequence tracks the farmhouse, well, and barn across disappearance and return, allowing comparison of object persistence and neighboring layout.The figure rows show methods and columns follow the video sequence.Source: G Additional Visual Comparisons · p. 23, Figure 9 · p. 24
• The town sequence tracks the workshop, chimney, crates, and loading barn as the camera passes behind a warehouse.Compare persistence across methods and video frames.Source: G Additional Visual Comparisons · p. 23, Figure 10 · p. 25
• The woodland turn-away and return sequence tracks the ranger cabin, log crossing, and wood rack across methods and video frames.The comparisons also show whether requested camera motion is completed.Source: G Additional Visual Comparisons · p. 23, Figure 11 · p. 26