acceptodds
Under review as a conference paper at ICLR 2027

GeoTrans: Geometry-Aware State Transport for Training-Free World Model Acceleration

Abstract

Diffusion-based world models generate controllable videos through iterative denoising, but repeated transformer evaluations over long spatiotemporal sequences make inference expensive. Training-free caching exploits temporal redundancy, yet temporal prediction alone does not capture the shared state evolution of geometrically corresponding tokens. Our analysis reveals that geometric correspondence organizes coherent state changes while substantial residual variation remains, motivating geometry-aware latent state transport. We propose GeoTrans, a training-free acceleration framework that determines when to refresh anchors, how to construct geometric transport, and what state update to accept. GeoTrans adaptively refreshes anchors using temporal drift and a horizon limit, decomposes proposed displacements into geometric shared and residual components, and estimates their gains online in closed form from completed anchor transitions. Geometry defines the transition structure, while anchor transitions determine the accepted update, requiring no additional transformer evaluations or gradient computation for gain estimation. Experiments on Voyager and Aether achieve and acceleration, respectively, with improved generation fidelity over WorldCache. Controlled ablations further isolate the benefit of geometric state transport.

Primary area:
generative models
Keywords:
World Models, Training-Free Acceleration, Geometry-Aware State Transport, Inference Acceleration, Latent State Prediction

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_pLw6OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a training-free acceleration method for diffusion-based world models. It combines adaptive anchor refresh with geometry-aware decomposition of temporally predicted state displacements into shared and residual components. Separate gains for these components are estimated online from completed full-anchor transitions using closed-form regularized regression, without additional transformer evaluations. Experiments report 4.25× acceleration on Voyager and 1.96× on Aether generation, with improved fidelity relative to WorldCache.

Critical strengths

The author provides a useful geometric extension to temporal state prediction. The method uses geometric correspondence to define shared and residual displacement components while retaining token-specific variation. Online calibration is computationally lightweight and uses only completed anchor transitions. Meanwhile, the controlled experiments show that geometry contributes beyond generic calibration. Table 4 compares operators under the same full-evaluation schedule while allowing each variant to evolve its own sampler trajectory. GeoTrans improves PSNR by 0.62 dB over global scalar calibration and 0.82 dB over shuffled correspondence.

Critical weaknesses
  1. The cross-model geometry construction is insufficiently specified. Appendix A.2 clearly describes supplied geometry for Voyager, but Appendix D.3 does not establish equally clearly how the fixed correspondence map is obtained for each Aether task. In particular, the reconstruction setting requires clarification of which depth and camera quantities are available before inference and which are prediction targets.

  2. The reliability of the incremental quality improvements is not quantified. Although the full-benchmark evaluation is a strength, several improvements over WorldCache are relatively small: +0.33 dB PSNR and +0.23/+0.16 WorldScore points on Voyager, and +0.17 dB PSNR on Aether. The paper does not report paired scene-level uncertainty or variability across generation seeds. Consequently, it is difficult to assess how consistently the quality advantage holds, particularly beyond the Voyager operator ablation.

Questions for authors
  1. For Aether generation and reconstruction, what exact depth and camera information is used to construct the fixed correspondence map, and when is it available?

  2. For LingBot, how are the correspondence map and gain updates constructed under the four-call chunk schedule?

  3. Were accelerated methods and full-step references evaluated using matched random seeds?

  4. Can you provide paired scene-level confidence intervals for the main quality differences, especially Voyager WorldScore and Aether PSNR, and state how many generation seeds were used?

5 comments
AuthorsOpenReview ↗

Response to Reviewer pLw6 (Part 1)

We thank the reviewer for their valuable comments and constructive feedback. We address each point below.


R3-W1: Cross-model geometry construction and availability of depth and camera information for Aether tasks.

Response to R3-W1:

  • Aether generation: Camera trajectories are supplied through raymap conditioning, while MoGe estimates depth from the conditioning image before denoising. The correspondence map remains fixed during sampling, with geometry acquisition included in runtime.
  • Video-depth reconstruction: MoGe estimates depth/intrinsics, while optical flow and RANSAC-PnP estimate poses from observed RGB frames. Maps are constructed before denoising and fixed per window, with all preprocessing costs included in runtime. Evaluated depth remains Aether’s prediction.
  • Camera-pose reconstruction: No external depth or pose estimates are used. Skipped-step raymaps are constrained to analytic camera-ray geometry (affine directions and a shared origin per frame) within the timed sampler. Evaluated poses remain Aether’s predictions.

No evaluation annotations, full-step references, or Aether denoising outputs are used for map construction.

Revisions. Appendix D.3 now clarifies geometry sources, availability, prediction targets, and runtime accounting for each task. Fixed correspondence maps apply to generation and video-depth; the scalar fallback below 10% correspondence support applies only to video-depth, while camera-pose uses analytic raymap constraints.


R3-W2: Statistical reliability of quality improvements over WorldCache across scenes and generation seeds.

Response to R3-W2:

For Voyager, we evaluate 20 paired scenes across seeds 0, 1, and 2, using identical inputs and initial noise. Both methods use 12 FULL evaluations, with matched denoising times (240.7 s WorldCache vs. 240.4 s GeoTrans). Unlike Table 1’s default WorldCache setting (3.65× vs. 4.25×), this comparison isolates fidelity at equal inference cost.

Table 1. Voyager quality differences (GeoTrans − WorldCache) with paired 95% CIs.

SeedΔPSNR ↑ (dB)ΔLPIPS ↓ΔWorldScore-Static ↑
0+2.25 [1.85, 2.72]−0.045 [−0.061, −0.032]+0.17 [−1.59, 2.01]
1+2.89 [2.19, 3.68]−0.050 [−0.065, −0.037]−0.83 [−2.65, 0.90]
2+2.74 [2.04, 3.57]−0.045 [−0.053, −0.036]+0.11 [−1.11, 1.27]
Pooled+2.62 [2.20, 3.08]−0.046 [−0.056, −0.038]−0.19 [−1.26, 0.84]

PSNR and LPIPS improvements are consistent across seeds, with CIs excluding zero, while WorldScore differences remain statistically inconclusive.

For Aether generation, 20 paired scenes with identical warm-up and FULL-evaluation positions yield +0.81 dB PSNR (95% CI [0.54, 1.10]).

For Aether reconstruction, we evaluate two seeds on MPI-Sintel using WorldCache’s 30-step protocol at comparable inference cost.

Table 2. Aether reconstruction results by seed with paired 95% CIs.

MetricMethodSeed 42Seed 43
AbsRel ↓WorldCache0.42100.3961
GeoTrans0.35940.3387
ATE ↓WorldCache0.19140.2584
GeoTrans0.17310.2137
Speedup ↑WorldCache3.32×3.32×
GeoTrans3.39×3.39×

GeoTrans reduces AbsRel by 14.6%/14.5% and ATE by 9.6%/17.3% for seeds 42/43. Paired 95% CIs for GeoTrans − WorldCache are [−0.087, −0.038]/[−0.083, −0.034] (AbsRel) and [−0.037, −0.003]/[−0.095, −0.008] (ATE). All intervals exclude zero. Speedups are relative to 30-step Aether.

Revisions. We added per-seed results, paired confidence intervals, and resampling procedures to the experimental appendix. The revised discussion distinguishes robust fidelity gains from smaller, statistically inconclusive WorldScore differences.

AuthorsOpenReview ↗

Response to Reviewer pLw6 (Part 2)

R3-Q1: Depth and camera information used for Aether correspondence maps and its availability.

Response to R3-Q1: Generation and video-depth reconstruction use fixed correspondence maps, while camera-pose reconstruction uses analytic raymap constraints.

Table 3. Geometry sources and availability for Aether tasks.

Task / protocolDepthCamera informationAvailability and cost
GenerationMoGe depth from input imageSupplied trajectory (raymap) and intrinsicsBefore denoising; 0.9 s/scene included in runtime
Reconstruction: video-depthMoGe depth from observed framesMoGe intrinsics; optical flow and RANSAC-PnP poses using MoGe 3D pointsBefore each window; all preprocessing costs included
Reconstruction: camera-poseNot usedAnalytic camera-ray constraints; no external pose estimateWithin the timed sampler

Correspondences use integer pixel centers, visibility checks, and a 10% relative depth-consistency threshold. Projected patches are grouped by reference-grid cells across latent frames. Maps remain fixed within each sampling window or rollout. Video-depth falls back to scalar calibration when retained groups cover less than 10% of tokens.

For camera-pose, skipped-step raymaps are constrained to affine ray directions and a shared ray origin per frame. Evaluated depth and poses remain Aether's predictions; no evaluation annotations, full-step references, or denoising outputs are used to construct maps.

Revisions. Appendix D.3 now details task-specific geometry sources, construction, availability, and runtime accounting.


R3-Q2: Correspondence map construction and gain updates under LingBot’s four-call chunk schedule.

Response to R3-Q2: The four-call schedule separates first-chunk calibration from subsequent prediction:

  1. Geometry construction: Before denoising, MoGe depth and supplied camera poses/intrinsics define the correspondence map. Poses are interpolated to latent frames, and intrinsics are scaled to 464×832. Correspondences index the 29×52 token grid of the intermediate DiT block-stack state before the prediction head. The map remains fixed throughout the rollout.
  2. First-chunk calibration: All four calls execute exactly. The first three states (H0,H1,H2H_0,H_1,H_2) form a proposal for the fourth before its exact output is observed. Shared/residual components are separated by the chunk-local projector, and the fourth exact state provides the sole calibration target for fitting two gains (Eq. 20).
  3. Subsequent chunks: The first three calls remain exact; the fourth block-stack state is predicted using fixed gains. No further calibration occurs. Prediction history and grouping remain chunk-local, while the native prediction head and KV updates always execute fully.

Across four chunks, both methods replace three block-stack evaluations: 3/20 = 15% including KV updates, or 3/16 = 18.75% of denoising calls. These represent call-count reductions.

We evaluate per-chunk fidelity against matched full rollouts over the first 49 of 61 generated frames.

Table 4. Per-chunk PSNR with first-chunk-only calibration.

ChunkFramesWorldCache ↑GeoTrans ↑ (gain [95% CI])
0 (calibration)1–1344.3844.39 (+0.00 [−0.01, 0.01])
114–2913.3315.05 (+1.72 [1.57, 1.89])
230–4511.9013.75 (+1.85 [1.36, 2.38])
346–4910.8012.32 (+1.51 [0.99, 2.09])

Gains denote GeoTrans − WorldCache, with paired scene-level 95% CIs. GeoTrans maintains positive PSNR gains across all three prediction chunks.

Revisions. Appendix D.3 now clarifies geometry construction, the pre-head interface, calibration ordering, and fixed-gain reuse. Table 2 specifies the call-count denominator.

AuthorsOpenReview ↗

Response to Reviewer pLw6 (Part 3)

R3-Q3: Matched random seeds between accelerated methods and full-step references.

Response to R3-Q3: All full-step and accelerated evaluations use matched random seeds, initial diffusion noise, and conditioning-latent sampling. Fidelity is measured against the corresponding full-step output for each scene–seed pair.

The seed configurations are:

  • Voyager: Seed 0 for baseline reproduction (40 single-scene and 10 three-scene prompts); seeds 0, 1, 2 for variability analysis.
  • Aether: Seed 42 for generation; seeds 42, 43 for reconstruction.
  • LingBot: Seed 42 for chunked generation.

For closed-loop rollouts, each method conditions subsequent scenes on its own generated final frame, preserving autoregressive error propagation.

Revisions. Appendices D.2–D.3 now document seed settings and conditioning-latent sampling, while Section 5.1 clarifies the pairing protocol.


R3-Q4: Paired scene-level confidence intervals and generation seed counts for Voyager WorldScore and Aether PSNR.

Response to R3-Q4: We provide 95% paired bootstrap CIs using 10,000 scene-level resamples. For Voyager's three-seed evaluation, differences are averaged across seeds per scene before resampling. The original benchmark generation results use one seed per prompt; revised Aether reconstruction results average two seeds.

Table 6. Voyager WorldScore-Static at matched denoising cost.

EvaluationScenes × SeedsWorldCacheGeoTransΔ [95% CI]
Seeds 0, 1, 220 × 362.4362.24−0.19 [−1.26, 0.84]
Seed 040 × 164.4265.21+0.79 [−0.56, 2.17]

Δ denotes GeoTrans − WorldCache. Voyager PSNR/LPIPS CIs are provided in Table 1 of Response to R3-W2.

Table 7. Aether generation: 20 scenes, seed 42.

MetricWorldCacheGeoTransΔ [95% CI]
PSNR ↑34.1534.97+0.81 [0.54, 1.10]
LPIPS ↓0.02750.0192−0.0083 [−0.0095, −0.0072]
SSIM ↑0.96810.9747+0.0067 [0.0052, 0.0082]

Both methods share identical inputs, noise, warm-up (10 steps), and FULL-evaluation positions (30/50 steps). Two-seed Aether reconstruction results appear in Table 2 of Response to R3-W2.

Results. Voyager WorldScore CIs include zero in both panels, while PSNR/LPIPS improvements are consistent across three seeds. Aether generation also shows positive fidelity improvements with CIs excluding zero. We have revised the discussion to distinguish these statistically supported fidelity gains from the inconclusive WorldScore differences.

Revisions. The experimental appendix now documents paired resampling, seed counts, and confidence intervals for each evaluation.


We sincerely thank the reviewer for the valuable feedback, which has helped us improve the clarity and rigor of our paper. We hope our responses, additional experiments, and revisions have adequately addressed your concerns. We remain active throughout the rebuttal period and welcome any further questions or discussion. If our clarifications have resolved your concerns, we would greatly appreciate your consideration of a more favorable assessment.

Reviewer_XZKzOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GeoTrans reframes training-free caching in diffusion world models as a geometry-aware state transport problem. The framework decomposes skipped-step denoising into three coupled decisions: (i) when to refresh an anchor, using adaptive shadow-drift monitoring; (ii) how to construct state transport, via depth-based geometric correspondence groups that split displacements into shared and residual components; and (iii) what state update to accept, through closed-loop calibration of per-component gains from completed anchor transitions. Evaluated on Voyager-13B and Aether-5B, GeoTrans reports up to 4.25x acceleration with improved generation fidelity over WorldCache.

Critical strengths
  1. Novel conceptual framework: reframing caching as geometry-aware state transport, with a shared/residual displacement decomposition, unifies prior scattered methods (WorldCache's curvature caching, HERO's hierarchical extrapolation) under one structured perspective and provides a research agenda for future work.
  2. Broad and multi-metric experimental validation: evaluation spans two model families (Voyager, Aether) and covers image generation, 3D reconstruction settings, with metrics including WorldScore, PSNR/SSIM/LPIPS, FVD, and geometric reconstruction accuracy (AbsRel, ATE, RPE).
  3. Rigorous ablation design: controlled closed-loop ablations (Table 3-4) isolate the contribution of each component (refresh policy, geometric correspondence, gain calibration) by keeping all other factors fixed.
  4. Training-free and model-agnostic in principle: no additional transformer evaluations, gradient computation, or offline fitting of the pretrained model are required; anchor states and gain estimation are scheduled online from completed anchor transitions.
Critical weaknesses
  1. Robustness of the core geometric assumption is untested: the entire method rests on depth-projected correspondences, but depth estimation error, occlusions, and textureless regions (sky, white walls) are never analyzed, so the reader cannot judge failure modes or practical deployment limits.
  2. Baseline fairness is questionable: WorldCache's hyperparameters use its original paper settings while GeoTrans's threshold is tuned on pilot scenes, so the 4.25x vs. 3.65x speedup gap may overstate GeoTrans's advantage and undermine the headline quantitative claim.
  3. Error accumulation over long rollouts is not discussed: gain estimation from past anchor transitions can propagate errors along a trajectory, but no analysis shows whether errors diverge over long horizons, which matters for a method whose selling point is reduced inference latency.
  4. The Aether 3D reconstruction experiment is weakly integrated with the main WHEN/HOW/WHAT pipeline: the connection between geometric transport decisions and downstream reconstruction quality is not thoroughly analyzed, making the contribution of geometry-aware transport to reconstruction gains unclear.
Questions for authors
  1. Could you inject controlled noise or masking into the depth input to test how PSNR and WorldScore degrade when the correspondence map becomes unreliable? This would clarify the method's practical applicability on scenes where depth estimation is known to be poor.
  2. On longer rollouts (e.g., 100+ frames), does accumulated skip horizon correlate with growing generation error? Do you observe any drift divergence beyond a certain trajectory length?
  3. Did you re-tune WorldCache's hyperparameters on the same 200-scene WorldScore subset used for GeoTrans? If you force both methods to the same operating-point budget (e.g., same number of FULL evaluations per sequence), how much of the speedup gap remains?
  4. The Aether experiments evaluate 3D reconstruction metrics (AbsRel, ATE, RPE). How sensitive are these downstream geometry metrics to the geometric transport decisions versus simple temporal caching? Could the reconstruction gains be attributed primarily to the anchor refresh policy rather than the geometric correspondence?
6 comments
AuthorsOpenReview ↗

Response to Reviewer XZKz (Part 1)

We thank the reviewer for their valuable comments and constructive feedback. We address each point below.


R2-W1: Robustness to depth estimation errors, occlusions, and textureless regions.

Response to R2-W1: We have added a controlled depth-stress evaluation and clarified how GeoTrans handles unsupported correspondences.

GeoTrans does not require every token to have a valid geometric match: projected matches that fail visibility or depth-consistency checks are excluded from the shared component, and tokens without a retained match remain in the residual component, calibrated by β⊥\beta_{\perp}. When no usable geometric support remains, the update reduces to scalar-calibrated temporal transport.

We evaluate 20 WorldScore prompts, perturbing only the depth used for correspondence construction while keeping model conditioning, random inputs, and FULL-evaluation positions fixed. Log-depth noise tests estimation error; a contiguous mask provides a controlled proxy for missing geometry in occluded or textureless regions; complete masking removes geometric support entirely.

Table 1. Sensitivity to depth perturbations used for correspondence construction.

Depth conditionMean geometric supportΔPSNR ↑ (dB) [95% CI]ΔLPIPS ↓ [95% CI]
Unperturbed depth0.7414——
Log-depth noise, σ=0.2\sigma=0.20.3219+0.00 [−0.06, +0.06]+0.0011 [−0.0002, +0.0024]
50% contiguous mask0.2913+0.02 [−0.02, +0.08]+0.0000 [−0.0020, +0.0020]
Complete mask0.0000+0.01 [−0.04, +0.07]+0.0008 [−0.0005, +0.0021]

Support is the fraction of tokens assigned to retained correspondence groups. Δ denotes perturbed − unperturbed; brackets contain paired 95% confidence intervals.

Despite substantial reductions in geometric support, the mean quality changes remain close to zero, and all paired confidence intervals include zero. These results support robustness under the tested depth-noise and missing-depth conditions; systematically wrong camera trajectories are outside the scope of these tests. The benefit of correct correspondences is assessed separately by the controlled geometry ablations in Table 4 of the paper.

We have expanded Appendix A.2 to explain correspondence filtering, residual handling, and scalar fallback. We have also added a “Robustness to Imperfect Geometry” subsection in the experimental appendix documenting the evaluation panel, perturbations, fixed FULL-evaluation schedule, and paired quality comparisons.


R2-W2: Baseline fairness in hyperparameter tuning and speedup comparisons.

Response to R2-W2: GeoTrans’s threshold τ\tau is calibrated once from shadow-drift statistics on 10 pilot scenes: the median of the per-scene peak drift under horizon-only refresh. No generation-quality metric or comparison against WorldCache enters this calibration (Appendix A.4), and the resulting threshold is held fixed in the main Voyager evaluation.

Beyond WorldCache’s default operating point, Fig. 6 varies its maximum skip horizon, with its other published hyperparameters retained, and evaluates both methods on the same 200-sample subset. Table 2 lists three WorldCache configurations and two representative GeoTrans points; all points are shown in Fig. 6.

Table 2. Representative speed–quality operating points from Fig. 6 on the 200-sample subset.

Method / configurationLatency (s) ↓PSNR ↑LPIPS ↓SSIM ↑
WorldCache, nmax⁡=6n_{\max}=6290.95623.6760.13930.8045
WorldCache, nmax⁡=7n_{\max}=7282.49623.6070.13810.7868
WorldCache, nmax⁡=8n_{\max}=8275.22022.9060.14420.7793
GeoTrans247.85823.2850.14140.7840
GeoTrans268.67823.8070.13240.8045

Both GeoTrans points have lower latency than every WorldCache configuration shown. The higher-fidelity GeoTrans point also achieves higher PSNR and lower LPIPS than all three WorldCache configurations, while matching or exceeding their SSIM. The faster GeoTrans point improves all three fidelity metrics over the fastest WorldCache configuration shown.

The 4.25× and 3.65× speedups in the paper’s Table 1 describe individual operating points; the sweep shows that GeoTrans’s speed–quality advantage persists across WorldCache’s skip horizons. When both methods are forced to the same budget of 12 FULL evaluations per sequence, GeoTrans still improves PSNR by 2.62 dB (Table 8 of Response to R2-Q3).

We have clarified the threshold-calibration procedure in Appendix A.4 and identified the configurations associated with the plotted operating points and the shared evaluation subset in the Fig. 6 caption, and expanded the experimental discussion to compare multiple baseline operating points.

AuthorsOpenReview ↗

Response to Reviewer XZKz (Part 2)

R2-W3: Error accumulation and stability over long rollouts.

Response to R2-W3: Two mechanisms limit error accumulation: (1) FULL evaluations refresh predictor anchors, with extrapolation capped by nmax⁡n_{\max}; (2) shared/residual gains are updated via EMA (ρ=0.5\rho=0.5), clipped to [−1,1][-1,1], and bounded by Eq. (26) to prevent amplification.

We evaluate 10 paired prompts over 145-frame, three-scene closed-loop rollouts using WorldScore's protocol. Each scene conditions on the previous generated final frame, preserving error propagation. Mean denoising times are 240.558 s (GeoTrans) and 240.900 s (WorldCache), differing by only 0.14%.

Table 3. Scene-wise fidelity over 145-frame closed-loop rollouts at matched latency.

MetricScene 1Scene 2Scene 3
PSNR ↑ WorldCache20.132115.008412.4272
PSNR ↑ GeoTrans22.540916.486313.5514
ΔPSNR [95% CI]+2.4087 [1.7290, 3.1654]+1.4779 [0.8760, 2.0970]+1.1242 [0.7647, 1.5097]
LPIPS ↓ WorldCache0.19890.40880.5307
LPIPS ↓ GeoTrans0.15400.35520.4833

ΔPSNR denotes GeoTrans − WorldCache; paired 95% CIs resample complete rollouts. Final-scene LPIPS improves by 0.0474 (95% CI [0.0343, 0.0615]).

Although fidelity decreases for both methods, GeoTrans maintains positive PSNR gains with CIs excluding zero through the final scene, supporting its effectiveness over the tested horizon.

Revisions. We clarified anchor refresh and bounded calibration around Eq. (26) and added closed-loop results with paired CIs to the appendix.


R2-W4: Contribution of geometry-aware transport to Aether 3D reconstruction quality.

Response to R2-W4: We have clarified the WHEN/HOW/WHAT integration and added two-seed reconstruction results at comparable inference cost. We follow WorldCache's 30-step Aether protocol on MPI-Sintel, using uncached Dense30 as the speedup reference.

  • WHEN: Horizon/drift triggers control FULL evaluations. GeoTrans/WorldCache use 8/9 FULL evaluations per 30-step window for video-depth and 9/9 for camera-pose.
  • HOW: Transport operates on x^0\hat{x}_0. Skipped steps preserve observed RGB and enforce analytic camera-ray constraints. Video-depth additionally uses MoGe and optical flow/PnP correspondences; camera-pose requires no external geometry.
  • WHAT: Gains are calibrated online from completed FULL transitions, using correspondence-based shared/residual grouping for video-depth. FULL outputs remain unchanged.

Both methods share inputs, checkpoint, scheduler, evaluation protocol, and four-step warm-up. WorldCache's threshold is selected on held-out sequences using denoising time only.

Table 4. Primary Aether reconstruction metrics by seed at comparable inference cost.

MetricMethod / StatisticSeed 42Seed 43
AbsRel ↓WorldCache0.42100.3961
GeoTrans0.35940.3387
Improvement−14.6%−14.5%
95% CI[−0.087, −0.038][−0.083, −0.034]
ATE ↓WorldCache0.19140.2584
GeoTrans0.17310.2137
Improvement−9.6%−17.3%
95% CI[−0.037, −0.003][−0.095, −0.008]
Depth speedup ↑WorldCache3.32×3.32×
GeoTrans3.39×3.39×

Percentages are relative to WorldCache; brackets denote paired sequence-level 95% CIs for GeoTrans − WorldCache.

Table 5. Additional reconstruction metrics for seeds 42 and 43.

MetricWorldCache (42)GeoTrans (42)WorldCache (43)GeoTrans (43)
δ<1.25\delta<1.25 ↑0.4390.4990.4400.528
δ<1.252\delta<1.25^2 ↑0.6470.7120.6570.738
SqRel ↓1.5911.0921.6801.160
RMSE ↓3.5122.9143.5472.978
RPE-trans ↓0.07300.06400.06680.0597
RPE-rot ↓0.7990.6960.7620.664

For camera-pose, speedups are 3.32×/3.30× (WorldCache/GeoTrans). Video-depth timing includes geometry acquisition (MoGe, optical flow/PnP, grouping); the camera-pose constraint is computed inside the timed sampler.

GeoTrans improves all reconstruction metrics across both seeds, with paired CIs excluding zero for AbsRel and ATE. The fixed-refresh comparison in Table 9 (Response to R2-Q4) further shows that these gains persist under identical FULL-evaluation positions, supporting the contribution of skipped-step transport beyond refresh placement.

Revisions. Section 5 and Appendix D.3 now clarify the WHEN/HOW/WHAT integration, transport interface, geometric constraints, and calibration. Revised Table 2 reports two-seed averages, with detailed results and timing definitions in the appendix.

AuthorsOpenReview ↗

Response to Reviewer XZKz (Part 3)

R2-Q1: Robustness to depth noise, masking, and unreliable correspondence maps.

Response to R2-Q1: We have performed this controlled test on 20 WorldScore prompts. To isolate sensitivity to unreliable correspondences, we perturb only the depth copy used for correspondence construction; model conditioning, random inputs, and GeoTrans’s FULL-evaluation positions remain identical to the unperturbed run.

We test log-depth noise with σ=0.2\sigma=0.2, a 50% contiguous mask, and complete masking. WorldCache is configured to match GeoTrans’s denoising latency: 240.8 s versus 240.5 s per clip. Since WorldCache does not consume the perturbed correspondence-depth copy, its matched control outputs are reused across conditions.

Table 6. Fidelity and WorldScore under controlled correspondence-depth perturbations.

Depth conditionPSNR ↑ WorldCachePSNR ↑ GeoTrans (gain [95% CI])LPIPS ↓ WorldCacheLPIPS ↓ GeoTransWorldScore-Static ↑ GeoTrans
Unperturbed depth21.1023.35 (+2.25 [1.85, 2.72])0.2050.16164.83
Log-depth noise, σ=0.2\sigma=0.221.1023.35 (+2.25 [1.83, 2.72])0.2050.16264.85
50% contiguous mask21.1023.37 (+2.27 [1.86, 2.74])0.2050.16164.81
Complete mask21.1023.36 (+2.26 [1.84, 2.74])0.2050.16264.85

PSNR gains denote GeoTrans − WorldCache; brackets contain paired 95% confidence intervals. Geometric-support statistics and quality changes relative to unperturbed depth are reported in Table 1 of Response to R2-W1.

The observed PSNR and LPIPS changes remain small across the tested perturbations, while GeoTrans retains its fidelity advantage over latency-matched WorldCache. The WorldScore-Static difference between complete masking and unperturbed depth is +0.02, with a paired 95% CI of [−1.17, 1.12], consistent with zero change. These results support robustness to noisy or missing correspondence-depth information under the tested conditions.

We have included the perturbation protocol, matched-latency baseline configuration, WorldScore endpoint evaluation, and paired confidence intervals in the “Robustness to Imperfect Geometry” subsection of the experimental appendix.


R2-Q2: Error accumulation and drift divergence over long rollouts.

Response to R2-Q2: We evaluated WorldScore’s native three-scene rollouts: 10 paired prompts spanning 145 unique frames. Each method continues from its own generated final frame, without teacher forcing. WorldCache is evaluated at matched denoising latency.

Skip horizon. The total number of predicted calls increases with rollout length, while the local skip age—the number of skipped evaluations since the latest FULL anchor—remains bounded by nmax⁡n_{\max}, because each FULL evaluation refreshes that anchor. GeoTrans’s realized FULL count and maximum skip age remain unchanged across the three scenes:

Table 7. GeoTrans’s realized skip statistics over the three-scene rollout.

SceneCumulative unique framesFULL evaluations per sceneMaximum skip age
149126
297126
3145126

Each scene uses a 50-step denoising schedule. Thus, the experiment measures rollout error as the autoregressive trajectory grows while keeping the local extrapolation horizon bounded.

Error over rollout length. Scene-wise PSNR, LPIPS, and paired confidence intervals are reported in Table 3 of Response to R2-W3. Fidelity to the full-step reference decreases for both methods as the rollout proceeds, while GeoTrans retains its advantage in every scene, including the final scene extending beyond 100 frames, with positive paired PSNR confidence intervals throughout.

Within this 145-frame native three-scene evaluation, the scene-level results show no distinct late-rollout collapse of GeoTrans relative to WorldCache.

We have expanded the “Long-Rollout Evaluation” subsection in the experimental appendix to report the realized FULL counts and maximum skip ages alongside scene-wise fidelity, and to distinguish the bounded local skip horizon from cumulative rollout length.

Reviewer_aUQfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GeoTrans, a training-free method for accelerating diffusion world models. It uses camera and depth correspondences to split temporally predicted state displacements into a shared component and a residual component. Separate gains are estimated online from completed anchor transitions, while a drift threshold and a horizon limit determine when to perform full model evaluations. Experiments on Voyager and Aether report improved fidelity and lower runtime than WorldCache, with additional results on LingBot. The paper also includes closed-loop ablations that compare different update operators under a common full-evaluation schedule.

Critical strengths
  1. The geometric decomposition is well motivated. Corresponding tokens can share part of their update while retaining token-specific changes through the residual. The two gains have a simple closed-form solution and use observations from already scheduled model evaluations. The derivations in Appendix C appear correct, and the paper appropriately distinguishes its update-magnitude bound from a guarantee on prediction error or rollout stability. This provides a useful way to incorporate geometry into temporal prediction at low additional cost.

  2. Table 4 provides a meaningful control for the contribution of geometric grouping. The variants share full-evaluation times and the temporal prediction rule, but each runs its own sampler trajectory. The scalar-gain baseline tests whether uniform calibration is sufficient, while shuffled correspondences preserve the group-size distribution and the two-gain parameterization. The reported improvements of 0.62 dB over scalar calibration and 0.82 dB over shuffled correspondences support the choice of geometric groups.

Critical weaknesses
  1. The basis of the main comparison is unclear. Table 1 states that results are averaged over the full 2,000-scene WorldScore benchmark, and Section 5.1 states that all experiments use an A100. However, the Voyager, TaylorSeer, HiCache, TeaCache, EasyCache, and HERO rows match the corresponding rows in WorldCache’s Table 1 across all reported columns. For example, TaylorSeer has the same PSNR of 18.32 and latency of 1195.2 seconds, while EasyCache has the same PSNR of 21.76 and latency of 294.5 seconds. WorldCache’s public version, arXiv:2603.06331v2, describes an A800 and an evaluation set of 40 single-scene prompts and 10 three-scene prompts in Appendix C.1–C.2.

    Neither the experimental section nor Appendix D.2 explains whether these baseline numbers were reproduced or taken from the earlier paper. The numerical agreement alone does not establish that the comparison is invalid, but the difference in the stated protocols needs an explanation. Without knowing which results were obtained under common conditions, I cannot assess the main speed–quality advantage reliably.

  2. The cross-model implementations leave important questions about geometry availability and execution. Algorithm 1 constructs fixed correspondences before denoising. Appendix D.3 says that Aether’s depth and raymap components provide the geometry, but does not explain where these quantities come from or when they become available for each task. This is particularly important for reconstruction, where depth and camera parameters are themselves predicted outputs. D.3 also describes three modality-specific token blocks produced by patch embedding, whereas Aether’s original formulation concatenates modalities along the channel dimension. An additional split or rearrangement at the output interface could explain this, but it is not specified.

    The LingBot adaptation is also incomplete: with chunk-local history, three exact calls followed by a predicted fourth call, it is unclear which subsequent full transition supplies the target for updating the calibration gains. These details determine whether the method can operate with the stated inputs and computational budget, so they are necessary to interpret Table 2.

Questions for authors
  1. Which rows in Table 1 were rerun, and which, if any, were taken from prior work? Please explain the agreement with WorldCache’s published values and provide the configurations or records establishing a common evaluation set, hardware setup, and timing scope for the full model, WorldCache, and GeoTrans. Does “2,000-scene WorldScore” refer to the static split, and how was the Dynamic column computed?

  2. For Aether generation and reconstruction separately, what depth and camera information is used to build the correspondence map, when is it available, and is its acquisition included in the reported runtime? Please specify the tensor layout and interface where the modality-specific operators are applied. Also, how many denoising steps are used for reconstruction, and how does the reported 3.05× speedup relate to the four-step reconstruction setting described in the original Aether paper?

  3. For LingBot, are the calibration statistics retained across chunks, or are the gains fixed after the first chunk? If they continue to change, which full evaluations provide their targets? Please also clarify the geometry source and the denominator used for the reported 15% block-evaluation reduction.

6 comments
AuthorsOpenReview ↗

Response to Reviewer aUQf (Part 1 )

We thank the reviewer for their valuable comments and constructive feedback. We address each point in detail below.


R1-W1: The basis of the main comparison is unclear.

Response to R1-W1: Thank you for pointing out the ambiguity in the provenance and evaluation protocol of Table 1. In the submitted table, Vanilla (full-step Voyager), WorldCache, and GeoTrans were measured by us, whereas the TaylorSeer, HiCache, HERO, TeaCache, and EasyCache results were taken from WorldCache’s published Table 1, which explains the numerical agreement of those five rows. The revision annotates these sources explicitly.

The comparison underlying our main claim, WorldCache versus GeoTrans, uses rows measured by us under one protocol: the same 2,000 static-split prompts, a single A100 80GB, the same seed, and the same timing definition. The quoted rows serve as reference points. Both WorldCache’s Voyager experiments and ours use single-GPU inference, and the A100 and A800 80GB variants have the same specified compute throughput and memory bandwidth (Table 1, from NVIDIA’s A100 specifications and A800 datasheet). NVLink provides GPU-to-GPU communication and is not used in single-GPU runs.

Table 1. Relevant specifications of the NVIDIA A100 80GB and A800 80GB.

SpecificationA100 80GBA800 80GB
FP32 throughput19.5 TFLOPS19.5 TFLOPS
TF32 Tensor Core throughput156 TFLOPS156 TFLOPS
FP16/BF16 Tensor Core throughput312 TFLOPS312 TFLOPS
GPU memory80GB HBM2e80GB HBM2e
NVLink bandwidth600 GB/s400 GB/s

Compute figures are dense peak rates without sparsity. Memory bandwidth is compared between corresponding form factors.

We further provide our Vanilla results on the 2,000-scene WorldScore static split (Table 2) and our remeasurements of the five quoted baselines under WorldCache’s sampling protocol of 40 single-scene and 10 three-scene prompts (Table 3). The revision distinguishes these two evaluation sets.

Table 2. WorldCache-reported Vanilla results and our 2,000-scene static-split reference results.

SourceWorldScore-Static ↑WorldScore-Dynamic ↑Latency (s) ↓
WorldCache paper: reported Vanilla66.2846.401053.7
Our Vanilla evaluation66.2546.381051.2

The Dynamic column follows WorldCache’s reporting convention for aggregation on static-split outputs.

The five baseline reproductions use a single NVIDIA A100 80GB GPU, with fidelity measured against the corresponding Vanilla outputs. We report synchronized denoising latency, including cache and CPU-offloading transfers within the denoising loop, while excluding model loading and final video decoding/saving.

Table 3. WorldCache-reported results and our remeasurements of the five baselines, including CPU-offloading configurations.

MethodReported PSNR ↑Our PSNR ↑Reported LPIPS ↓Our LPIPS ↓Reported lat (s) ↓Our lat (s) ↓CPU offloading: reportedCPU offloading: ours
TaylorSeer18.3218.040.2930.2961195.21204.1YesYes
HiCache18.5618.690.2810.2791100.11082.3YesYes
HERO17.7118.160.3150.2971100.01013.2YesYes
TeaCache16.2517.910.3720.264311.5401.6——
EasyCache21.7622.450.2080.174294.5331.5——

“Reported” denotes values from WorldCache’s Table 1; “Our” denotes our remeasurements. Latency is measured per scene clip.

Our Vanilla scores and runtime closely match the published reference, and the baseline remeasurements preserve the broad speed–fidelity trends underlying the original comparison.

We have revised Table 1 and its caption to explicitly identify self-measured and quoted results and state their respective evaluation scopes. We have also expanded Section 5.1 and Appendix D.2 to distinguish the 2,000-scene Vanilla reference set from the baseline reproduction panel, include the additional reference and baseline comparisons, and specify the hardware, CPU-offloading configurations, and timing definition.

AuthorsOpenReview ↗

Response to Reviewer aUQf (Part 2)

R1-W2: The cross-model implementations leave important questions about geometry availability and execution.

Response to R1-W2: We have clarified geometry acquisition and execution in the revision. Correspondence maps use task-provided or input-estimated geometry before denoising, without backbone outputs, evaluation annotations, or full-step references.

  • Aether generation. Camera trajectories are provided by raymap conditioning, while depth is estimated from the conditioning image using MoGe before denoising.
  • Aether reconstruction. For video-depth, MoGe estimates depth and intrinsics, while optical flow and RANSAC-PnP estimate poses from observed RGB frames. A fixed correspondence map is constructed before denoising, with all preprocessing costs included in runtime. For camera-pose, no external estimates are used; skipped-step raymaps are constrained by analytic camera-ray geometry, and depth and poses remain Aether's predictions.
  • LingBot. Supplied camera parameters and MoGe-estimated depth define a fixed correspondence map before denoising, while prediction history and correspondence grouping remain chunk-local.

Configurations and results are detailed below. Parenthesized quality values indicate GeoTrans − WorldCache differences, while percentages indicate relative changes over WorldCache.

Aether image-to-world generation. We generate 41 frames (480×720) with 50 steps and dynamic CFG (scale 3). Fidelity is measured against uncached outputs using identical inputs and noise. Both methods share a 10-step warm-up and 30 identical FULL evaluations. Denoising takes 99.1 s (WorldCache) and 99.2 s (GeoTrans), with 0.9 s geometry overhead included in GeoTrans's reported speedup.

Table 4. Aether image-to-world generation results.

MetricWorldCacheGeoTrans
PSNR ↑34.1534.97 (+0.81)
LPIPS ↓0.02750.0192 (−0.0083)
SSIM ↑0.96810.9747 (+0.0067)

Aether RGB-video reconstruction. We use Aether's 30-step sliding-window inference on MPI-Sintel training sequences. Depth is evaluated by valid-pixel-weighted AbsRel and δ<1.25\delta<1.25; pose by ATE and translational/rotational RPE after official Sim(3) alignment.

GeoTrans transports x^0\hat{x}_0 to decode depth/raymap. Skipped steps preserve observed RGB and enforce analytic raymap constraints, with input-video correspondences for video-depth. FULL outputs remain unchanged.

Both methods use a four-step warm-up; WorldCache's threshold is tuned on held-out sequences by denoising time. GeoTrans/WorldCache use 8/9 FULL evaluations for video-depth and 9/9 for camera-pose. Speedups are relative to Dense30 and include geometry preprocessing or in-sampler constraints. Results average seeds 42 and 43.

Table 5. Aether reconstruction at matched inference budgets.

MetricWorldCacheGeoTrans
AbsRel ↓0.40850.3490 (−14.6%)
δ<1.25 ↑0.43940.5133
ATE ↓0.22490.1934 (−14.0%)
RPE-trans ↓0.06990.0619
RPE-rot ↓0.78020.6801
Depth speedup ↑3.32×3.39×

Camera-pose speedups are 3.32×/3.30× (WorldCache/GeoTrans). For seeds 42/43, AbsRel improves by 14.6%/14.5% and ATE by 9.6%/17.3%, with paired 95% confidence intervals excluding zero. GeoTrans improves all reconstruction metrics at comparable latency.

LingBot few-step causal generation. We generate 61 frames across four causal chunks along WorldScore trajectories, evaluating the first 49 against full rollouts with identical seeds and conditioning. All four first-chunk calls are exact; the fourth-call proposal uses the preceding three states and is calibrated only against the fourth exact output. Gains remain fixed thereafter.

Table 6. LingBot few-step causal generation results.

MetricWorldCacheGeoTrans
PSNR ↑20.9022.19 (+1.29)
LPIPS ↓0.63290.5783 (−0.055)
SSIM ↑0.38920.4625 (+0.073)

Both methods replace 3/20 calls including KV updates (15%), or 3/16 denoising calls (18.75%). The native prediction head and KV updates execute fully and are excluded from calibration.

Output interface. Aether retains its native 96-channel input (56 noisy + 16 RGB + 24 raymap) and joint patch embedding. Its 56-channel output (B,Tlat,56,Hlat,Wlat)(B,T_{\mathrm{lat}},56,H_{\mathrm{lat}},W_{\mathrm{lat}}) is partitioned into RGB/depth/raymap (16/16/24). Generation transports RGB/depth displacements before the original scheduler update; reconstruction uses the same partition in x^0\hat{x}_0 space.

Revisions. Algorithm 1 clarifies map construction and validity; Appendix D.3 details geometry, transport interfaces, channel layouts, and LingBot calibration; Table 2 specifies configurations, Dense30 references, and runtime accounting.

AuthorsOpenReview ↗

Response to Reviewer aUQf (Part 3)

R1-Q1: Clarification of Table 1 reproducibility, evaluation consistency, and WorldScore static/dynamic protocols.

Response to R1-Q1: As clarified in Response to R1-W1, Vanilla (full-step Voyager), WorldCache, and GeoTrans in the submitted Table 1 were measured by us. The TaylorSeer, HiCache, HERO, TeaCache, and EasyCache rows were taken from WorldCache’s published results, which explains their exact numerical agreement. Please see Tables 1–3 in Response to R1-W1 for the hardware specifications, Vanilla reference comparison, and five-baseline reproductions, including CPU-offloading configurations.

For the Full/WorldCache/GeoTrans comparisons, we have expanded Appendix D.2 to identify the evaluation manifests, corresponding full-step reference outputs, generation configurations, and timing records. The evaluation scope, hardware, and timing definition are specified for each comparison.

These measurements are distinguished from the additional five-baseline reproduction panel, which follows WorldCache’s sampling protocol of 40 single-scene and 10 three-scene prompts. The reproductions use the corresponding outputs from our Vanilla reference set, seed 0 with seeded conditioning sampling, one NVIDIA A100 80GB GPU per job, 50 denoising steps, and CFG 6.0. Latency is synchronized denoising time, including cache and CPU-offloading transfers where applicable, while excluding model loading and final decoding/saving.

The 2,000 scenes refer to WorldScore’s static split, which our Vanilla reference set covers. The Dynamic column uses the official WorldScore aggregation on the same static-split outputs, following WorldCache’s reporting convention. The three motion-dependent components—motion accuracy, motion magnitude, and motion smoothness—are not evaluated on this split and enter the aggregate as zero. We have explicitly documented this aggregation convention in the revised Table 1 caption and Appendix D.2.


R1-Q2: Aether geometry acquisition, modality-specific implementation, and reconstruction speedup settings.

Response to R1-Q2: Geometry for generation. Camera trajectories are provided by raymap conditioning, while depth is estimated from the conditioning image using MoGe (as in Voyager) before denoising. Both geometry acquisition and correspondence construction are included in GeoTrans’s runtime.

Geometry for reconstruction. The input consists only of observed RGB video, with two protocols:

  • Video-depth: MoGe estimates depth/intrinsics, while bidirectional optical flow and RANSAC-PnP estimate relative poses using MoGe 3D points. Correspondence maps are constructed before denoising and fixed per window. All preprocessing costs are included in runtime.
  • Camera-pose: No external geometry is used. Skipped-step raymaps are constrained to analytic camera-ray geometry (affine ray directions and a shared origin per frame) within the timed sampler.

Evaluated depth and poses remain Aether’s predictions. Model loading is excluded consistently across methods.

Tensor layout and interface. Aether retains its native 96-channel input (56 noisy state + 16 RGB + 24 raymap) and joint patch embedding. The 56-channel output (B,Tlat,56,Hlat,Wlat)(B,T_{\mathrm{lat}},56,H_{\mathrm{lat}},W_{\mathrm{lat}}) is partitioned into RGB/depth/raymap (16/16/24) after unpatchification. Generation transports RGB/depth displacements before scheduler updates; reconstruction transports x^0\hat{x}_0, preserving observed RGB and analytic raymap constraints on skipped steps. FULL outputs remain unchanged.

Reconstruction steps and speedup. We follow WorldCache’s 30-step reconstruction protocol (Appendix C.1) using Aether’s native sliding-window inference on MPI-Sintel. Dense30 is uncached Aether with the same checkpoint and scheduler, evaluating all 30 steps.

The reported 3.05× speedup is relative to Dense30, independent of Aether’s native four-step setting. Revised Table 2 recomputes speedups using our measured Dense30 baseline under matched inference budgets, including geometry preprocessing. The 30-step schedule enables meaningful step-skipping comparisons, unlike the four-step setting.

Revisions. Appendix D.3 now specifies geometry sources, acquisition timing, channel layout, and transport interfaces. Table 2 clarifies the 30-step protocol, Dense30 reference, and runtime accounting.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal

Dear Area Chair,

GeoTrans is a training-free method that accelerates diffusion world models by using geometry and online calibration to predict skipped denoising steps. We evaluate it on Voyager, Aether, and LingBot across generation and reconstruction tasks.

Main reviewer feedback. Reviewer aUQf found the geometric decomposition “well motivated.” Reviewer XZKz highlighted the “Novel conceptual framework” and “Broad and multi-metric experimental validation.” Reviewer pLw6 noted that “geometry contributes beyond generic calibration.”

Discussion and ratings. We posted point-by-point responses and followed up individually to invite further questions. The table summarizes the discussion as of this writing.

ReviewerInitial ratingCurrent ratingConfidenceStatus
aUQf2: Weak rejection2: Weak rejection3Author response on 15 Nov; follow-up on 18 Nov; no reviewer reply to date
XZKz3: Weak acceptance3: Weak acceptance4Author response on 15 Nov; follow-up on 18 Nov; no reviewer reply to date
pLw63: Weak acceptance3: Weak acceptance3Author response on 15 Nov; follow-up on 18 Nov; no reviewer reply to date

Reviewer concerns and our responses

Reviewer aUQf

  • W1 / Q1 — Baseline provenance and comparability. We clarified which Table 1 results were measured by us and which were quoted from WorldCache, documented the shared protocol for the main Voyager comparison, and remeasured the five quoted baselines on a separate panel. We also clarified how WorldScore-Dynamic is computed on the static split. See R1-W1 and R1-Q1.
  • W2 / Q2 — Aether geometry and runtime. We documented when geometry becomes available for each task, how the model output is split for transport, and which geometry costs are included in timing. We clarified that the submitted reconstruction speedup was measured against uncached 30-step Aether, rather than the native four-step setting. See R1-W2 and R1-Q2.
  • W2 / Q3 — LingBot calibration and call counts. We identified the fourth exact call of the first chunk as the sole calibration target. The gains are then fixed, and the new per-chunk evaluation shows that the fidelity advantage persists in later chunks. The reported 15% reduction counts three of 20 block-stack calls, including KV updates. See R1-W2 and R1-Q3.

Reviewer XZKz

  • W1 / Q1 — Robustness to imperfect depth. We added controlled depth-noise and masking experiments. Across the tested conditions, mean PSNR changed by at most 0.03 dB, and WorldScore-Static showed little change. See R2-W1 and R2-Q1.
  • W2 / Q3 — Baseline tuning and matched budgets. We clarified GeoTrans’s drift-based threshold selection, swept WorldCache’s skip horizon, and added an equal-budget comparison. With 12 full evaluations per sequence and nearly identical denoising latency, GeoTrans gained 2.62 dB PSNR. See R2-W2 and R2-Q3.
  • W3 / Q2 — Error accumulation over longer rollouts. We added 145-frame closed-loop rollouts at matched latency. Fidelity declined for both methods, while GeoTrans retained its advantage through the final scene (+1.12 dB PSNR). See R2-W3 and R2-Q2.
  • W4 / Q4 — What drives the reconstruction gains? We linked the reconstruction implementation to the main method and added two-seed results at comparable cost. In a separate control with identical refresh positions, GeoTrans still reduced depth and pose errors, showing that refresh placement alone does not explain the gains on that panel. The control supports the overall reconstruction design without attributing the gains to a single component. See R2-W4 and R2-Q4.

Reviewer pLw6

  • W1 / Q1–Q2 — Geometry sources across tasks. We distinguished supplied cameras and input-image depth estimates for Aether generation, correspondences estimated from observed RGB for video-depth reconstruction, and in-sampler ray constraints for camera-pose reconstruction. We also documented how LingBot builds its map before denoising and uses it with chunk-local prediction history. See R3-W1 and R3-Q1–Q2.
  • W2 / Q3–Q4 — Matched seeds and statistical reliability. We documented how conditioning and initial noise are matched to the full-step references and added paired confidence intervals and seed-level results. The fidelity gains are consistent across three Voyager seeds, and the controlled Aether generation comparison shows a positive PSNR gain. The Voyager WorldScore intervals include zero, and the revision distinguishes that finding from the fidelity results. See R3-W2 and R3-Q3–Q4.

We believe the additional experiments and clarifications address each of the concerns raised by the reviewers and make GeoTrans’s contributions and scope clearer. Thank you again for your time and consideration.

Kind regards,

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.