INSPATIO-WORLD: A Real-Time 4D World Simulator via Spatiotemporal Autoregressive Modeling
1. Paper overview
In one sentence: Reference-anchored autoregressive diffusion and geometric reprojection support controllable video roaming, while dual-teacher distillation targets visual fidelity and generated-region memory remains incomplete. e-architecturee-cachee-geometrye-jdmde-worldscoree-long-videoe-rerender-protocole-evaluation-scope
| At a glance | What to know |
|---|---|
| Research problem | Source description Long camera trajectories expose accumulated scene drift, weak control and synthetic-looking appearance. The target is interactive novel-view video generation from monocular observations, with user-supplied camera motion. The paper motivates embodied intelligence and driving as downstream possibilities; it does not evaluate an acting robot or driving policy. e-probleme-identitye-evaluation-scope |
| Core mechanism | Source description STAR couples a persistent reference anchor, recent generated history and explicit geometric guidance inside autoregressive diffusion. e-architecturee-cachee-geometry |
| A key reported result | WorldScore interactive world generation: 68.72; 81.51 Dynamic Overall ↑; Camera control ↑. 1.3B model; reported WorldScore benchmark comparison, with real-time/interactive rows identified. TeleWorld: 66.73; 76.58. Noninteractive FantasyWorld-1.0: 71.39; 81.45. Best dynamic overall among the listed interactive methods and best listed camera score, but not best overall dynamic score across all methods. TeleWorld also leads in 3D consistency and static overall. e-worldscore |
| Reading caution | Source description The authors acknowledge incomplete long-term texture memory for newly generated regions and unresolved consistency of moving elements during wide-angle, omnidirectional roaming. Structural point-cloud memory does not preserve every generated appearance detail. e-limitations |
Core contributions
- Source description
STAR couples a persistent reference anchor, recent generated history and explicit geometric guidance inside autoregressive diffusion. e-architecturee-cachee-geometry
- Author claim
JDMD shares student weights across synthetic-data camera-controlled rerendering and text-only distillation from a real-video foundation-model teacher. The authors argue this preserves control while improving appearance. e-jdmd
Figure 2. Camera instructions become geometric guidance; the reference and latest generation provide memory. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start in the blue panel: user input determines a target camera trajectory, while reference content supplies depth and the reference latent. Warping produces guidance and a mask, joined through the explicitly labeled channel-concatenation path. The purple cache contains the reference latent and previous generated block. Follow the upper feedback arrow from the denoised block back toward that cache, and the other branch toward accumulated latents and VAE decoding. The dashed lower panel belongs to training: motion and perceptual teacher/critic signals feed JDMD gradients into the generator. Sections 3.2–3.3 explain these pathways more precisely than the compact drawing. e-architecturee-cachee-geometrye-initializatione-jdmde-evaluation-scope
What it supports. The design gives the generator both an appearance anchor and viewpoint-aligned structure. JDMD adds a second training distribution to the same student, rather than requiring two teachers to generate each interactive frame. The architecture explains the intended information flow; the benchmark comparisons do not isolate how much each path contributes.
Where the evidence stops. The figure calls m an occlusion mask, while Equations (2)–(3) describe a valid-pixel mask; bit polarity is unspecified. Read the plus-shaped input node through the channel-concatenation label, not as a proven arithmetic sum. Teacher/critic connections are schematic; Equation (4) distinguishes their score functions.
2. Motivation
2.1 The problem and the proposed response
Long camera trajectories expose accumulated scene drift, weak control and synthetic-looking appearance. The target is interactive novel-view video generation from monocular observations, with user-supplied camera motion. The paper motivates embodied intelligence and driving as downstream possibilities; it does not evaluate an acting robot or driving policy. e-probleme-identitye-evaluation-scope
2.2 What this reading follows
Imagine moving the camera away from a reference video, then returning to a room that the generator must still recognize. INSPATIO-WORLD addresses this with two kinds of guidance: cached reference and recent-frame information, and a depth-based rendering of the requested viewpoint. Its training also separates learning camera compliance from recovering realistic appearance through two teacher distributions. Read the architecture before the tables: the reported interactive system uses a 1.3B model, whereas the long-video and rerendering comparisons use 14B. The results show strong generation quality, but neither component ablations nor successful embodied action execution are supplied. e-architecturee-cachee-geometrye-jdmde-worldscoree-long-videoe-rerender-protocole-evaluation-scope
3. Research context
We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.
| Catalog dimension | Recorded classification |
|---|---|
| Major category | Foundational work |
| Architecture | Not applicable |
| Prediction paradigm | Not applicable |
| Quadrant | Not applicable |
This table preserves the labels recorded at reading time. The current major category is Benchmarks & simulators. View the current classification.
3.1 Evidence-based assessment
Supports the recorded classification
The recorded foundational neural-world-simulator classification fits reference-conditioned scene generation with explicit geometry. The WAM action-prediction quadrants are not applicable: camera instructions are supplied externally and converted to poses, while the learned output is video. Shared JDMD weights are architecture evidence for a generator, not evidence of joint future/action prediction or an inverse-dynamics controller. e-architecturee-geometrye-jdmd
This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.
4. Problem formulation
4.1 Inputs and outputs
| Inputs | Outputs |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 From a requested viewpoint to a conditioned prediction
Section 3.2 gives the user instruction a geometric route into the generator. A relative camera update changes the accumulated pose; feedforward reconstruction supplies depth and camera intrinsics; reprojection then aligns reference content with that pose. The mask tells the generator which projected regions are valid, so an empty region need not be interpreted as genuinely black texture. Meanwhile, the reference latent and preceding generated block enter through the cache. The former anchors appearance in observed content, while the latter supplies recent temporal context. Geometry channels are zero-padded for historical blocks, so an old viewpoint constraint is not carried forward as though it were the current instruction. The resulting latent block is predicted and decoded into imagery; no action policy is learned by this sequence. e-architecturee-cachee-geometrye-initialization
5.2 Why the two teachers appear during training
Camera-controlled rerendering needs synchronized views that are scarce in real video, making synthetic training useful but potentially restrictive in appearance. JDMD therefore alternates two tasks over shared student weights. In V2V iterations, reference and geometric conditions are supervised through the synthetic-data motion teacher. In T2V iterations, text alone conditions the student and the original Wan-T2V teacher supplies the real-video distribution target. Equation (4) distinguishes teacher scores from student-tracking fake scores; Equation (5) balances the resulting visual and control objectives. Before distillation, causal rehearsal establishes the conditioning behavior. During distribution-loss training, gradient-free rollout followed by per-chunk recomputation allows cache construction and denoising to receive gradients within a chunk. The paper asserts appearance transfer without loss of control, but supplies no isolated JDMD comparison. e-jdmde-initializatione-cachee-evaluation-scope
5.3 Separate a convincing system result from a mechanism proof
Reader interpretation: the tables support a useful camera-conditioned video generator more directly than they support every causal explanation in the text. RE10K-Long improves all four reported metrics, whereas rerendering exposes a quality/control tradeoff: NeoVerse retains lower Blender pose errors. WorldScore adds an interactive comparison, but its 1.3B configuration differs from the 14B models in the other tasks. Figure 3 further uses a parameter–step proxy rather than measured end-to-end latency. These distinctions prevent the 24 FPS headline and the strongest quality numbers from becoming one unsupported operating point. They also motivate separate mechanism checks. A stable reference anchor can explain consistency for observed content, yet the authors explicitly acknowledge incomplete texture memory for newly generated regions and unresolved wide-angle dynamic consistency. e-worldscoree-long-videoe-rerenderinge-rerender-protocole-efficiencye-implementatione-cachee-limitations
5.4 Training and inference
During training
Initialization rehearses chunk-wise, multi-step autoregression using ground-truth data or teacher ODE trajectories. It establishes the heterogeneous conditions before few-step distillation. e-initialization
Chunk-wise backpropagation first generates the full sequence without gradients to obtain the distribution-loss signal, then recomputes cache construction and denoising chunk by chunk with gradients. The stated differentiability guarantee is within each chunk; full cross-chunk gradient equivalence is not demonstrated. e-cache
JDMD alternates V2V and text-only T2V tasks. The frozen motion teacher represents synthetic-data control; the frozen original Wan-T2V teacher represents real-video appearance. Student-tracking fake-score networks supply the other side of distribution matching. This is training supervision, not two-teacher inference. e-jdmd
Training uses Wan2.1 and Self-Forcing, internet videos including RealEstate10K, UE renders and ReCamMaster. Teacher and initialization learning rates are 2×10⁻⁵; distillation uses 4×10⁻⁶ for the student and 8×10⁻⁷ for the fake-score network. e-implementation
During inference
For each new user instruction, update pose, prepare geometry and reference/history conditions, denoise the current chunk, and carry its output forward. Tiny-VAE replaces Wan-VAE and torch.compile accelerates streaming; the authors acknowledge a small quality cost from the VAE substitution. e-architecturee-geometrye-implementation
5.5 Implementation flow
- Turn interaction into image-aligned geometry
User operations produce a six-degree-of-freedom relative pose update accumulated into the current camera pose. Feedforward reconstruction supplies depth and intrinsics; reference features are reprojected into the requested view. A binary validity mask distinguishes unavailable pixels from black texture. Optional reconstruction of generated frames expands a point-cloud memory. e-geometry
- Keep a reference anchor and recent history
The cache combines the corresponding reference latent with the preceding generated block in a sliding window. Fixed starting position indices for current, reference and historical blocks constrain RoPE extrapolation. Constant memory is claimed for the KV cache, not for every optional scene representation. e-cachee-geometry
- Denoise the next chunk
The DiT predicts blocks of K consecutive frames from noisy latents conditioned on cache content and warped features plus mask. Geometry channels apply only to the current block; historical geometry channels are zero-padded. Generated blocks feed subsequent history and are VAE-decoded into video. e-architecturee-initialization
6. Experiments & results
INSPATIO-WORLD turns a reference video into camera-controlled generated video. Its STAR generator combines a reference/history cache with depth-based reprojection; JDMD distills motion and appearance guidance into shared student weights. Benchmark gains support this generative simulator, while the small-model speed claim and larger-model quality results require separate interpretation.
The supplied paper contains full-system benchmark comparisons and qualitative examples, but no component-removal ablation, memory-retention curve, JDMD-only comparison or gradient-interference diagnostic. Accordingly, this edition includes the original efficiency comparison and long-video examples as diagnostics of overall behavior, without presenting either as a mechanism-isolating ablation. The source provides no ablation page to crop. e-evaluation-scopee-efficiencye-long-examples
6.1 Read the original evidence
Table 1. The interactive ranking is favorable, but different methods lead different WorldScore dimensions. Original paper, p. 10 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. First use the Real-time/Interactive column to define the comparison group. The final two rows are TeleWorld and INSPATIO-WORLD, both marked Yes. Read Dynamic Overall and Camera together: these describe aggregate dynamic performance and camera following, rather than a single universal quality score. Then move across the same rows to 3D Consistency, Object, and Static Overall to see which tradeoffs disappear in a headline ranking. All displayed arrows point upward. Bold values mark the best reported value in a column and underlining marks second place. Section 4.2 identifies the INSPATIO-WORLD entry as the 1.3B model. e-worldscoree-evaluation-scopee-identity
What it supports. INSPATIO-WORLD scores 68.72 in Dynamic Overall and 81.51 in Camera, versus TeleWorld’s 66.73 and 76.58. FantasyWorld-1.0 has a higher overall dynamic score, 71.39, but is marked noninteractive. TeleWorld’s 87.35 in 3D Consistency exceeds INSPATIO-WORLD’s 84.18, so the table supports a selective advantage.
Where the evidence stops. These are full-system comparisons without reported uncertainty or matched component ablations. Being first among the displayed interactive rows does not establish dominance on every metric or validate the downstream robot and driving applications pictured elsewhere.
Figure 3. The plotted efficiency advantage uses a parameter–step proxy, not measured latency. Original paper, p. 10 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read upward for a larger WorldScore-Dynamic value and leftward for a smaller parameter-count-times-step-count product. The numeric horizontal ticks are logarithmic; the far-right Non-public category is not another measured compute value. INSPATIO-WORLD is the red point, separated toward the upper left from many listed baselines. Compare its vertical position with the exact dynamic scores in Table 1 rather than extracting extra precision from marker placement. The shaded band and dashed line have no explanatory definition in the caption, so they should not be treated as confidence intervals or a documented fitted scaling law. e-efficiencye-worldscoree-implementatione-long-videoe-rerender-protocol
What it supports. The plot illustrates a favorable score-versus-proxy position for INSPATIO-WORLD at the reported 68.72 dynamic score. It is consistent with the authors’ efficiency motivation. The separate implementation paragraph supplies the actual 1.3B throughput claims: 24 FPS on an unspecified H-series GPU and 10 FPS on RTX 4090.
Where the evidence stops. Parameter count multiplied by denoising steps omits resolution, geometry preparation, decoder work and hardware effects. The non-public points are not quantitative compute measurements. Neither this plot nor the FPS paragraph establishes the latency of the 14B configurations used in later tables.
Table 2. The 14B long-video experiment improves both distribution metrics and camera errors. Original paper, p. 11 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each column downward before comparing across metrics: FID concerns generated image distributions, FVD concerns video distributions, and Rot/Trans summarize camera-trajectory error. Lower is better in all four columns. The bold bottom row is INSPATIO-WORLD; the underlined LingBot-World row is the strongest listed comparator throughout. The surrounding experimental text specifies 100 randomly selected RE10K sequences longer than 150 frames and says the 14B model is used to match LingBot-World. This matters because Table 1 and the advertised streaming rate refer to the smaller model. The four numbers are not interchangeable units or a combined score. e-long-videoe-metricse-worldscoree-implementatione-limitations
What it supports. INSPATIO-WORLD reports FID 42.68 and FVD 100.55, versus LingBot-World’s 64.84 and 173.02. Its Rot/Trans values are 2.8762/0.1398 versus 11.981/0.2064. This is the paper’s clearest joint improvement in distribution quality and camera control, within the described long-video evaluation.
Where the evidence stops. The paper does not supply uncertainty, exact sequence IDs, or camera-error units and implementation details. Sequences exceeding 150 frames are evidence for this sampled horizon, not indefinite scene persistence or the stability of newly invented content after arbitrary revisits.
Figure 4. Intermediate and final views help distinguish camera movement from merely preserving a plausible room. Original paper, p. 11 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Treat the two room scenes as separate examples. In each group, the left image is the source; subsequent columns show ground truth and four methods. The upper row is an intermediate frame and the lower row is the final frame, according to the caption. Follow stable landmarks such as the fireplace, sofa, windows and television across rows, then compare their placement with the ground-truth column. A plausible-looking room can still represent the wrong camera trajectory. HY-WorldPlay’s views remain relatively similar here, whereas other baseline columns show changed framing or structure. INSPATIO-WORLD follows the displayed ground-truth view progression more closely. e-long-examplese-long-videoe-limitations
What it supports. These selected examples make the long-video metrics easier to interpret: control failure can appear as insufficient viewpoint change, wrong framing or altered scene layout. The displayed INSPATIO-WORLD frames better preserve the requested view progression, while Table 2 supplies the aggregate comparison over the stated evaluation sample.
Where the evidence stops. Two selected image sequences are illustrative evidence, not a failure-rate estimate or proof of artifact-free navigation. Still frames cannot establish continuous motion quality, and these indoor examples do not resolve the dynamic omnidirectional-memory limitations acknowledged in Section 5.1.
Table 3. Rerendering quality improves, while the most accurate camera result depends on the dataset and metric. Original paper, p. 12 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Keep the OpenVid and Blender blocks separate. OpenVid lacks target ground-truth videos, so the paper reports six VBench dimensions and their arithmetic mean, Overall, together with camera errors. Blender has target videos and therefore includes FID and FVD. VBench arrows point upward; camera and distribution errors point downward. Compare INSPATIO-WORLD with NeoVerse across an entire block rather than stopping at bold entries. Sections 4.4 on pages 12–13 specify 240 OpenVid cases from 40 videos and six trajectories, 100 Blender samples, and the 14B INSPATIO-WORLD model. The table contains no uncertainty estimates. e-rerenderinge-rerender-protocole-implementatione-metrics
What it supports. On Blender, FID/FVD fall to 44.46/110.11 from NeoVerse’s 103.23/230.87, but NeoVerse has lower Rot/Trans. On OpenVid, the mean rises from 0.8486 to 0.8507 and translation error falls from 0.1340 to 0.1240; rotation error slightly worsens. Improved aggregate appearance does not imply the best result in every column.
Where the evidence stops. OpenVid’s no-target protocol cannot support target-video distribution comparisons. Several individual VBench scores favor baselines, and the small mean difference has no reported uncertainty. These 14B quality results should not be presented as measurements of the 1.3B real-time system.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| WorldScore interactive world generation 1.3B model; reported WorldScore benchmark comparison, with real-time/interactive rows identified. | 68.72; 81.51 Dynamic Overall ↑; Camera control ↑ | TeleWorld: 66.73; 76.58. Noninteractive FantasyWorld-1.0: 71.39; 81.45. Best dynamic overall among the listed interactive methods and best listed camera score, but not best overall dynamic score across all methods. TeleWorld also leads in 3D consistency and static overall. e-worldscore |
| RE10K-Long image-to-video generation 14B model; 100 randomly selected RE10K sequences exceeding 150 frames; the authors state identical inputs and size matching with LingBot-World. | 42.68; 100.55; 2.8762; 0.1398 FID ↓; FVD ↓; Rot ↓; Trans ↓ | LingBot-World: 64.84; 173.02; 11.981; 0.2064. All four metrics improve in this sample. Pose-error units and uncertainty are unspecified; this is not a 1.3B real-time quality measurement. e-long-videoe-metrics |
| Blender camera-controlled video rerendering 14B model; 100 synthetic samples with trajectories and ground-truth target videos. | 44.46; 110.11; 1.2386; 0.0667 FID ↓; FVD ↓; Rot ↓; Trans ↓ | NeoVerse: 103.23; 230.87; 1.2148; 0.0636. Distribution metrics improve substantially, while NeoVerse retains lower camera errors. These outcomes establish a tradeoff rather than uniform dominance. e-rerenderinge-rerender-protocol |
| OpenVid camera-controlled video rerendering 14B model; 240 samples from 40 videos paired with six trajectories; no ground-truth target videos. | 0.8507; 1.6000; 0.1240 VBench six-metric mean ↑; Rot ↓; Trans ↓ | NeoVerse: 0.8486; 1.5780; 0.1340. A small aggregate quality gain and lower translation error accompany slightly higher rotation error. No uncertainty is supplied, and several individual VBench metrics favor baselines. e-rerenderinge-rerender-protocol |
| Streaming generation throughput 1.3B model with Tiny-VAE and torch.compile. | Reported 24 FPS on an unspecified NVIDIA H-series GPU; 10 FPS on RTX 4090. Frames per second ↑ | No matched timing ablation or 14B throughput is supplied. Hardware-specific author-reported rates; missing resolution and end-to-end latency accounting prevent a complete interactive-performance comparison. e-implementation |
6.3 Ablations and diagnostic examples
No ablation or diagnostic visual is included in this edition. The source-specific visual limitations are recorded above.
7. Analysis & limitations
7.1 What the evidence leaves open
The authors acknowledge incomplete long-term texture memory for newly generated regions and unresolved consistency of moving elements during wide-angle, omnidirectional roaming. Structural point-cloud memory does not preserve every generated appearance detail. e-limitations
The experiments compare full systems without isolating the cache, geometry, initialization, recomputation or JDMD. Thus the tables do not establish their individual causal contributions, nor the claim that distinct task inputs prevent gradient interference. No robot-execution or closed-loop driving result is reported. e-evaluation-scopee-jdmd
7.2 Questions for discussion
- Which mechanism actually prevents return-view texture drift when the reference cannot show a newly generated region?
- Does JDMD improve real-video appearance at fixed control accuracy and matched training compute?
8. Reproducibility audit
8.1 Requirements and known gaps
Required assets include a Wan2.1 backbone, motion-teacher fine-tuning data, the real-video teacher, depth reconstruction, student-tracking score networks and the Tiny-VAE/compilation path. Training iterations, batch size, data mixture/counts, exact reconstruction checkpoint, chunk length, fixed position values, JDMD task schedule and λ_ctrl are unspecified. e-implementatione-cachee-geometrye-jdmd
A comparable evaluation needs the selected RE10K and rerendering samples, trajectories, camera-error implementation and metric settings. The text does not identify exact sample IDs, pose-error units, seeds or uncertainty procedures. Timing also needs the exact H-series device and configuration. The illustrated edition proposes controlled memory and JDMD comparisons. e-long-videoe-rerender-protocole-metricse-implementation
8.2 Proposed reproduction checks
The following checks are proposals motivated by the paper. They have not been run as part of this reading.
Check 1: Measure whether revisits need the reference anchor or the point-cloud map
Proposed, not performed: run a paired inference diagnostic on the same camera loops and random seeds with the reference-cache anchor intact versus masked, crossed with optional point-cloud memory enabled versus disabled. Keep the checkpoint, recent-history cache, position indices, denoising settings and reference-video input fixed. Include both previously observed surfaces and newly generated regions, using synthetic target views where available. Compare camera errors and return-view texture agreement only on mutually visible regions. Improvement confined to observed surfaces would support anchoring while exposing the stated new-region limitation. Masking is an inference intervention, so a training-matched follow-up would be needed to claim component necessity. e-cachee-geometrye-long-videoe-rerender-protocole-metricse-limitations
Check 2: Test whether the perceptual branch improves fidelity at fixed control
Proposed, not performed: compare JDMD with motion-teacher-only distillation from the same initialized student. Use one control matched for synthetic V2V exposure and a second matched for total training compute; keep evaluation trajectories, seeds, decoder and sampling settings fixed. Sweep the visual/control balance and report Blender FID/FVD together with Rot/Trans, then the six OpenVid dimensions separately. Bootstrap by source video so OpenVid’s six trajectories do not masquerade as independent source scenes. Fidelity gains that survive both budget controls without worse camera errors would support the claimed benefit; gains requiring degraded control would reveal a tradeoff. Report initialization and schedule details missing from the paper. e-jdmde-initializatione-implementatione-rerenderinge-rerender-protocole-metrics
8.3 Reading coverage
Visual audit: Visually inspected the title and author page, Figures 1–5, Tables 1–3, and all method, training, evaluation and limitation text on PDF pages 1–13. Cross-checked Figure 2 feedback and conditioning paths against Equations (2)–(5) and Sections 3.2–3.3; its mask naming and schematic score paths are explained in the architecture caution. Every final crop was inspected, including the corrected architecture crop with its complete legend and no caption. References on pages 14–20 were read through the six text chunks, but were not visually inspected. No appendix is present and separate supplements were not supplied. All retained scientific details and proposed-check premises are supported on the declared pages.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13. Appendix coverage: not present.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- Title and authors; Abstract (pp. 1–2)
- 1 Introduction (pp. 2–3)
- 2 Related Work (pp. 3–4)
- 3.1 Problem Formulation (p. 5)
- 3.2 Spatiotemporal Autoregressive Framework, including 3.2.1–3.2.3 (pp. 5–7)
- 3.3 Joint Distribution Matching Distillation (p. 8)
- 3.4 Implementation Details (pp. 8–9)
- 4.1 Experimental Setup (pp. 9–10)
- 4.2 WorldScore Benchmark (pp. 10–11)
- 4.3 Long-term Image-to-Video Generation (pp. 11–12)
- 4.4 Camera Controlled Generative Video Rerendering (pp. 12–13)
- 5 Discussion and Conclusions; 5.1 Limitation; 5.2 Future Work (p. 13)
- Acknowledgment and References (pp. 14–20)
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Separate supplemental material availability has not been fully verified.
- Identity/version: the inspected title and all 22 named contributors under InSpatio Team match the catalog. The supplied PDF identifies arXiv:2604.07209v2, 13 April 2026; the catalog submission date is 8 April 2026. Only v2 was supplied, so changes from v1 were not compared.
- All six supplied text chunks were read individually, including references through page 20. No appendix is present in this PDF. Separate supplements were not supplied.
- The extraction limitation above was addressed by inspecting PDF pages 1–13, all five figures and all three tables, plus every final crop. Reference pages 14–20 were read as text.
- Code, linked websites and models were not inspected; no experiments were reproduced. The title page supplies no institutional affiliations, so none are inferred.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e-identityPDF p. 1, title/author block, arXiv margin and Figure 1 caption
The title matches the supplied observed title. InSpatio Team credits 22 named contributors in alphabetical order. The margin identifies arXiv:2604.07209v2, 13 April 2026. Figure 1 describes downstream embodied/driving applications as potential.
Go to primary source ↓e-problemPDF p. 2, Abstract continuation and Section 1, three listed bottlenecks
The target is interactive generation from a reference video, addressing scene persistence, synthetic-to-real appearance shift and imprecise camera control.
Go to primary source ↓e-architecturePDF p. 5, Figure 2 and Section 3.1 Eq. (1); p. 6, Section 3.2 Eq. (2)
Chunk-wise diffusion conditions on history, reference and warped features with a mask. Figure 2 shows cache feedback, DiT denoising, VAE output, and a separate JDMD training panel.
Go to primary source ↓e-cachePDF p. 6, Section 3.2.1, all three paragraphs
Reference/current history cache, fixed position indices and gradient-free full rollout followed by differentiable per-chunk recomputation are described. Constant overhead refers to KV cache.
Go to primary source ↓e-geometryPDF p. 7, Section 3.2.2, Eq. (3) and final paragraph
User commands yield accumulated 6-DoF camera updates. Feedforward reconstruction and reprojection produce guidance and a validity mask; optional generated-video reconstruction expands a point-cloud map.
Go to primary source ↓e-initializationPDF p. 7, Section 3.2.3, final two paragraphs
Initialization rehearses autoregressive multi-step generation on ground truth or teacher ODE trajectories; history geometry channels are zero-padded.
Go to primary source ↓e-jdmdPDF p. 3, JDMD introduction; p. 8, Section 3.3, Eqs. (4)–(5)
Shared student weights alternate synthetic-control V2V and real-distribution T2V distillation under two frozen teachers. Fake scores track the student. The text asserts reduced interference but gives no isolating measurement or numeric λ_ctrl.
Go to primary source ↓e-implementationPDF pp. 8–9, Section 3.4, datasets, stage learning rates and acceleration paragraph
Wan2.1/Self-Forcing training uses internet, UE and ReCamMaster videos. Learning rates are 2×10⁻⁵, 4×10⁻⁶ and 8×10⁻⁷ as specified by stage/network. Tiny-VAE and torch.compile accompany reported 1.3B throughput of 24 FPS on H-series and 10 FPS on RTX 4090; iteration counts are not given.
Go to primary source ↓e-metricsPDF p. 9, Section 4.1, evaluation metric bullets
Rot/Trans compare generated camera motion to trajectories; FID/FVD compare image/video distributions; six VBench dimensions assess visual quality. Camera-error units and full implementation settings are absent.
Go to primary source ↓e-worldscorePDF p. 10, Table 1, INSPATIO-WORLD/TeleWorld/FantasyWorld-1.0 rows; Section 4.2
The evaluated model is 1.3B. Dynamic overall/camera scores are 68.72/81.51 versus 66.73/76.58 and 71.39/81.45. TeleWorld has higher 3D consistency (87.35 versus 84.18) and static overall (78.23 versus 75.81).
Go to primary source ↓e-efficiencyPDF p. 10, Figure 3, both axes and caption
The plot places WorldScore-Dynamic against model parameters in billions multiplied by inference steps, with a separate non-public region. The caption reports 68.72 for INSPATIO-WORLD; the plot is not a wall-clock benchmark.
Go to primary source ↓e-long-videoPDF p. 11, Table 2, all rows; Section 4.3, first two paragraphs
The comparison uses the 14B model and 100 randomly selected RE10K sequences exceeding 150 frames. INSPATIO-WORLD reports 42.68/100.55/2.8762/0.1398 for FID/FVD/Rot/Trans, versus LingBot-World 64.84/173.02/11.981/0.2064.
Go to primary source ↓e-long-examplesPDF p. 11, Figure 4 and caption; Section 4.3 continuing on p. 12
Two scenes show source, ground truth, and four methods at intermediate/final frames. The authors discuss pose drift, structural warping and static-frame degeneration in baseline outputs.
Go to primary source ↓e-rerenderingPDF p. 12, Table 3, OpenVid and Blender blocks, all method rows
OpenVid mean/Rot/Trans are 0.8507/1.6000/0.1240 versus NeoVerse 0.8486/1.5780/0.1340. Blender FID/FVD/Rot/Trans are 44.46/110.11/1.2386/0.0667 versus 103.23/230.87/1.2148/0.0636. The caption defines Overall as the six VBench metrics’ mean.
Go to primary source ↓e-rerender-protocolPDF pp. 12–13, Section 4.4, first paragraph across page boundary
Blender has 100 samples with trajectories and ground truth. OpenVid has 240 cases from 40 videos and six trajectories without target ground truth. The authors use 14B to match NeoVerse.
Go to primary source ↓e-limitationsPDF p. 13, Sections 5.1–5.2
New-region texture memory and dynamic 360-degree consistency remain limited; stronger semantic memory and physical priors are future work.
Go to primary source ↓e-evaluation-scopePDF pp. 9–13, Sections 4.1–4.4, Tables 1–3 and Figures 3–5
The experimental section contains WorldScore, long-video and rerendering full-system comparisons. It contains no component-removal ablation, measured gradient-interference study, or robot/driving execution experiment.
Go to primary source ↓8.5 Primary sources
INSPATIO-WORLD: A Real-Time 4D World Simulator via Spatiotemporal Autoregressive Modeling ↗
PDF · 9,396 extracted words
Source fingerprint
af0263459410f81f99159feda3d85d26d0ab83f645ab3eefbeccfaa90b85b061