acceptodds
Under review as a conference paper at ICLR 2027

Velox: Spatial Memory Makes World Models Efficient

Abstract

Video world models turn actions and camera motion into long visual rollouts, making video a natural interface for interactive simulation. Such interaction requires responsive generation, yet coherent rollouts must retain visual memory of what has already been generated. Chunked video diffusion models usually keep this context in growing KV caches, so later chunks become progressively slower. To address this bottleneck, we introduce Velox, a training-free attention router for region-level context selection using frustum slab correspondences from streaming geometry. Velox represents historical blocks as camera frustum slabs with depth intervals and obtains current-block depth through recency-priority multiwarp. For well-covered regions, frustum slab overlap softly modulates each layer’s model-derived block importance. Uncovered regions retain the modelderived route without geometric bias. The fixed token-block granularity matches FlashAttention-style tiled execution and reduces the attention term that grows with rollout length. Frustum slab construction runs asynchronously with denoising on the same H200. At chunk 48, Velox accelerates attention by 4.36× and denoising by 2.43×. On the full 158-case WBench navigation split, it remains within 0.15 API-free average points of dense LingBot-Fast-Cam. Visual results are available at https://anonymous.4open.science/w/velox/

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
World Model, Spatial Memory

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ebaPOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents Velox, a training-free sparse attention router for camera-controlled video world models. Historical token blocks are represented as frustum slabs derived from streaming geometry, while current-view geometry is estimated by warping historical observations into the target camera. Geometric overlap softly modulates model-derived block importance, with a fallback to model-derived routing when geometric coverage is insufficient. The method combines this routing mechanism with a hardware-aligned token layout, a sparse attention kernel, and asynchronous geometry processing. On LingBot-Fast-Cam, it reports 4.36x attention speedup and 2.43x denoising speedup at chunk 48, while remaining within 0.15 API-free average points of the dense baseline on the 158-case WBench navigation split.

Critical strengths
  1. The method offers a practical way to introduce region-level spatial correspondence into an existing generator without retraining. Softly combining geometric evidence with model-derived importance, rather than relying exclusively on geometry, is a sensible design for generated videos with imperfect reconstruction.

  2. The paper demonstrates an implemented acceleration pipeline rather than only theoretical sparsity. The sparse kernel, token layout, and asynchronous geometry processing yield substantial measured late-chunk speedups on a single H200, while maintaining aggregate benchmark quality close to the dense baseline.

Critical weaknesses
  1. The manuscript lacks qualitative comparisons of generated videos. At the time of my review, I could not find any visualizations at the anonymous link provided in the manuscript. Examples showing scene consistency, revisited regions, and camera following relative to dense attention and model-only sparse routing would make the results easier to assess. Visualizing selected historical blocks would also help validate the proposed geometric correspondences.

  2. Robustness to camera-following errors is insufficiently evaluated. Aligning the commanded trajectory with reconstructed history does not ensure that the generated chunk follows the intended viewpoint. Consequently, warped block correspondences may be incorrect even when geometric coverage is high. It would be helpful to evaluate whether soft fusion and local-context safeguards remain effective under such discrepancies.

Questions for authors

None

2 comments
AuthorsOpenReview ↗

Response to Reviewer ebaP (1/2)

Thank you for the review. You point to two clear gaps: the paper had no qualitative comparisons, and it did not test what happens when the generated camera moves away from the command. We address both in the revision (changes in blue). This part covers W1, and Part 2 covers W2.

W1: The manuscript lacks qualitative comparisons of generated videos, and no visualizations could be found at the anonymous link. Examples of scene consistency, revisited regions, and camera following relative to dense attention and model-only sparse routing would help. Visualizing selected historical blocks would also help validate the geometric correspondences.

Response to W1. First, we apologize. Our anonymous project page did not display its content when you checked it, because of a technical issue we missed at submission. The page now works and shows about 20 side-by-side videos of dense attention and Velox. Thank you for taking the time to look for it.

New qualitative comparison (Sec. 4.5, Figure 4, and App. A.2 of the manuscript). Dense attention and Velox start from the same input frame and case inputs, and we take frames at 25%, 50%, 75%, and 100% of the rollout instead of picking moments by hand. Among 20 reviewed pairs, which we did not choose by their results, Velox stays closer to the input scene in 7, the two are close in 7, and the result is mixed in 6. The paper shows all 7 better cases together with this count, so the advantage is not presented as universal. The clearest revisit is a walk through a bedroom that later returns toward the start: dense attention ends in an unrelated dark room, whereas Velox returns to the same bed cover, desk, and armchair. In other cases, dense attention fades a red figurine to pale pink and turns a bamboo path into a puddle, while Velox keeps both. Revisits like these are where geometry is designed to help, since the useful history is the place seen before.

Model-only routing (new App. A.5 and Figure 7 of the manuscript). We added Block-Sparse next to dense attention and Velox for three of these cases, with the same input frame and camera path. These are shortened rollouts of about 80 frames along the same camera path. At this short length, all three models stay close to the input scene, and the differences emerge late in long rollouts, as Figures 4 to 6 of the manuscript show for dense attention and Velox. For full-length rollouts, Table 1 in the manuscript compares Block-Sparse and Velox on the 158-case benchmark (one run each, no intervals), with Block-Sparse below Velox on spatial consistency (71.8 vs. 74.3), perspective consistency (75.7 vs. 77.0), and the Overall average (75.01 vs. 75.60).

Selected history blocks (new App. A.5 and Figure 8 of the manuscript). In the bedroom case, we take one bed block in chunk 8, which faces the same direction as the first chunk (3° apart) from closer in. We then draw the top 12 history blocks chosen by the model's ranking alone and after soft fusion with geometry:

Top 12 history blocksModel ranking onlyAfter soft fusion
Mean slab overlap with the query block0.1170.201
Blocks with positive overlap5 of 1210 of 12

Fusion moves the selection onto history blocks whose 3D slabs overlap the query block, so the geometric match points to the right region. Because the bed stays in view along this short path, its closest matching views are in recent chunks, and fusion selects exactly those overlapping views.

What is new. We did not make this clear enough in the submission. Frame-memory methods choose which past views to keep, and every region of the current chunk then sees the same views. Sparse video attention chooses blocks from attention scores alone. Velox works at the level of token blocks and combines both signals, so each region gets its own history, chosen by 3D overlap and the model's ranking together, without any training. For example, a road block and a patch of sky in the same chunk can attend to different past blocks.

AuthorsOpenReview ↗

Response to Reviewer ebaP (2/2)

W2: Robustness to camera-following errors is insufficiently evaluated. Aligning the commanded trajectory with reconstructed history does not ensure that the generated chunk follows the intended viewpoint, so warped block correspondences may be wrong even when geometric coverage is high. It would help to test whether soft fusion and the local-context safeguards remain effective under such errors.

Response to W2. Thank you for suggesting this test. High coverage only indicates that past geometry explains the commanded view. It does not guarantee that the model actually rendered that view. The revision discusses this failure (App. D.5, Sec. 6) and tests it in two ways (new App. A.4). Briefly, on all 158 real rollouts, the gap between Velox and dense attention does not grow with camera drift. Under a deliberate 15° geometry offset, the coverage test withdraws the geometric boost from more regions in two of three cases.

Test 1: drift in real rollouts (Table 4 of the manuscript). WBench scores how closely each generated video follows its commanded camera path (the navigation score, where lower means more drift). We split the 158 cases into three groups by this score, fixing the groups before comparing the models. If wrong matches under drift hurt Velox, the high-drift group should show the largest loss. Below are Velox minus dense attention, in points, with paired bootstrap 95% intervals:

Drift groupCasesMean of 8 automatic metrics
High52+0.41 [−0.94, +1.81]
Middle52+0.42 [−1.01, +1.93]
Low54+0.18 [−0.74, +1.15]

No group shows a clear loss on the eight-metric mean, and the high-drift group is not worse than the others. The only clear navigation-score loss is in the low-drift group (−1.68 [−3.18, −0.07], Table 4 in the manuscript). Perspective consistency, the metric with the largest overall loss (App. A.3), is also not worse in the high-drift group (Table 4 in the manuscript).

Test 2: a deliberate camera mismatch (Table 5 of the manuscript). We rotated the target cameras used by the geometry by 15° while the generator still received the original commands, which creates the kind of mismatch you describe. We ran three cases on shortened rollouts and recorded the fraction of current blocks that pass the coverage test and therefore receive the geometric boost:

CaseNo offset15° offset
A (returns to start)0.910.60
B0.390.29
C (bedroom, returns to start)0.830.88

In cases A and B, the warped views stop matching the stored geometry, so fewer blocks pass the test, and those blocks fall back to the model-only route instead of trusting the wrong overlap. In case C, coverage stays high and fusion runs on the shifted geometry, which matches the high-coverage, wrong-geometry situation you raise. Its frames look nearly the same as without the offset, although this is a visual check rather than a score.

Why the damage is limited by design. First, soft fusion multiplies the model's own block score by at most 1 + λ = 3 (Eq. 11), so a wrong match can raise a block's score at most threefold, and blocks that the model scores near zero stay out. Second, the previous chunk, the local neighbors, and the reference frame stay in every route (Eqs. 12 to 14), so local context does not depend on geometry. Third, the alignment anchors each target camera at the latest reconstructed pose (App. B.2), so the error is measured from the most recent frame rather than added up over the rollout, although drift in the reconstructed poses themselves can still build up. Together, these parts bound the effect of a wrong match.

Thank you again for pointing us to this test. Please let us know whether these results address your concern.

Reviewer_rXxDOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Velox is a training-free, inference-time attention router for chunked video world models whose visual KV cache grows with rollout length. Overall, the writing and experiments in this paper are not good enough. It is hard to understand the details of methodology and experimental design. Moreover, I have not seen any qualitative results in this work, which is important to evaluate the performance.

Given that the paper fundamentally falls short of the clarity and quality thresholds expected at ICLR, I recommend rejecting this paper. Additionally, such submissions generated by LLMs should be strongly discouraged.

Critical strengths

none

Critical weaknesses
  1. Poor writing. The main text read like an implementation log, e.g., "Earlier versions used a repeated chunk-local Gilbert order and a separate cross-chunk seam table. The current implementation instead…". Please move implementation history to the appendix.
  2. Unclear methodology design. The authors combine different technologies to accelerate the inference speed, but I cannot find the importance of different parts.
  3. Lack of solid experimental results. I have not found any qualitative results in this paper.
  4. Missing costs. There is no end-to-end wall-clock for a full rollout, including LingBot-map reconstruction, multi-warp, OBB scoring, and contention from the asynchronous worker on the shared GPU. There is also no report of GPU memory: the KV cache still grows, and a reconstruction model must be resident.
Questions for authors

is this draft auto-researched by LLMs?

2 comments
AuthorsOpenReview ↗

Response to Reviewer rXxD (1/2)

Thank you for the review and for the specific points on writing, component importance, qualitative results, and full costs. We have revised the paper to address each of them (changes in blue). We first answer your question, then W1 and W2 here, and W3 and W4 in Part 2.

Q1: Is this draft auto-researched by LLMs?

Response to Q1. No. The method, its implementation design, and the methodology text are our own work. We used AI tools to polish the language and as coding assistants. We take full responsibility for every part of the paper.

W1: Poor writing. The main text reads like an implementation log, for example "Earlier versions used a repeated chunk-local Gilbert order and a separate cross-chunk seam table. The current implementation instead …". Please move implementation history to the appendix.

Response to W1. Thank you for the exact quote. We agree that this sentence described how our code evolved rather than the method itself. We removed such implementation history from the main text, including phrases such as "earlier versions", "the current implementation", and "seam table". The Gilbert paragraph now states the construction in four sentences and points to App. D.1 for details. We also added a short roadmap at the start of the method: Sec. 3.1 sets up chunked generation and the block unit, Sec. 3.2 gives the geometric score and the coverage test, Sec. 3.3 builds the soft route, and Sec. 3.4 analyzes its cost.

W2: Unclear methodology design. The authors combine different technologies to accelerate inference, but the importance of the different parts is unclear.

Response to W2. We agree this was unclear. The parts of Velox play two roles. Four of them decide which history each region can see: the slab overlap with its coverage test, the soft fusion, the fallback for uncovered regions, and the local safeguards. The others, namely the Gilbert KV storage, the sparse kernel, and the background geometry worker, mainly make a given route run faster, and Figure 3 in the manuscript measures the speed of Velox as a whole. The question is therefore which of the route-deciding parts matter for quality, and the revision gives the following evidence.

  1. Geometry or no geometry (Table 1 of the manuscript). Block-Sparse uses the model's own ranking at the same budget and with the same safeguards, but without geometry. Adding geometry raises the averages by less than one point and spatial consistency by 2.5 points (71.8 → 74.3), and it also improves perspective consistency (75.7 → 77.0), where both sparse routes fall below dense attention. Block-Sparse has no per-case scores, so these gaps are reported without intervals.
  2. Against dense attention (new App. A.3). A paired test over all 158 cases finds no clear difference on 11 of 17 metrics. Velox is slightly better on two image-quality metrics and slightly worse on four others, the largest of which is perspective consistency (−4.7 points), as the paper now states.
  3. Wrong camera input (new App. A.4). We rotated the cameras used by the geometry on purpose in three shortened rollouts. In two cases, fewer regions pass the coverage test, and those regions fall back to the model-only route. In the third case, coverage stays high and fusion uses the shifted geometry.
AuthorsOpenReview ↗

Response to Reviewer rXxD (2/2)

W3: Lack of solid experimental results. There are no qualitative results in the paper.

Response to W3. Thank you for this point. The revision now adds qualitative results, and we are sorry that our anonymous project page showed nothing at review time. We missed a technical problem at submission, and the page now works.

New qualitative comparison (Sec. 4.5, Figure 4, and App. A.2 of the manuscript). Dense attention and Velox start from the same input frame and case inputs, and we take frames at 25%, 50%, 75%, and 100% of the rollout instead of picking moments by hand. Among 20 reviewed pairs, which we did not choose by their results, Velox stays closer to the input scene in 7, the two are close in 7, and the result is mixed in 6. The paper reports this count and shows all 7 better cases. In these cases, dense attention drifts late in the rollout. For example, in one case it leaves a bedroom and ends in an unrelated dark room, while Velox returns to the same bed, desk, and armchair. In another, dense attention slowly fades a red figurine to pale pink. The project page now shows about 20 side-by-side videos of dense attention and Velox.

Model-only routing and selected blocks (new App. A.5). Figure 7 of the manuscript adds Block-Sparse next to dense attention and Velox on shortened rollouts. At this short length all three stay close to the input scene, and the differences emerge late in long rollouts. For full-length rollouts, Table 1 in the manuscript compares them on the 158-case benchmark, where Block-Sparse is below Velox on spatial consistency (71.8 vs. 74.3) and the Overall average (75.01 vs. 75.60). Figure 8 of the manuscript shows which history blocks one query block selects. Soft fusion moves the selection onto history blocks whose 3D slabs overlap the query block, and the number of the 12 selected blocks that overlap it rises from 5 to 10.

W4: Missing costs. There is no end-to-end wall-clock for a full rollout, including LingBot-map reconstruction, multi-warp, OBB scoring, and contention from the asynchronous worker on the shared GPU. There is also no report of GPU memory, although the KV cache still grows and a reconstruction model must be resident.

Response to W4. Thank you for pointing out both issues. The revision now states the scope of each measurement (Sec. 4.2, App. B.4, and Sec. 6).

At chunk 48 on one H200, per-call attention time falls from 34.2 to 7.8 ms (4.36×), and the denoising of one chunk from 7.19 to 2.96 s (2.43×) (Sec. 4.2, Figure 3, and App. B.4 of the manuscript). That figure measures denoising while the geometry worker runs reconstruction, multiwarp, and OBB scoring on the same H200, so any slowdown from sharing the GPU is already inside the 2.43× speedup. Model loading, the warm-up of the first chunks, VAE decoding, and the final drain (the wait for the geometry worker to finish its last task) are not included.

Beyond these numbers (Sec. 6), Velox loads the reconstruction model, and its first chunks run dense attention until geometry is ready. These one-time costs are amortized as the rollout grows, which is where Velox gains most. On memory, we agree that the KV cache still grows. Velox is designed to cut compute rather than memory. It keeps every past block, which preserves the full history for revisits, and also the reconstruction model, so its peak memory is higher than that of dense attention.

We hope the revision and these answers address your concerns, and we are glad to clarify any remaining point.

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

Velox is a training-free router for selecting historical KV blocks in chunked video world models. It represents past regions as camera-frustum slabs with depth intervals and uses geometric overlap to softly modulate layer-specific model-derived importance. Coverage-aware fallback and temporal safeguards protect regions with unreliable geometry. The implementation aligns block selection with tiled sparse attention and runs geometry construction asynchronously on the same H200. On 48-chunk rollouts, the paper reports 4.36× attention and 2.43× chunk-denoising speedups at chunk 48, while retaining near-dense aggregate quality on all 158 cases of the WBench navigation split.

Critical strengths
  1. A coherent and well-motivated routing design. Spatial overlap is a useful signal for identifying relevant historical context, and the method combines it with the generator's layer-dependent importance rather than replacing that importance with a rigid geometric rule. The fallback for low-coverage regions and temporal safeguards address obvious failure modes. This makes the proposal technically convincing as a practical method for context selection.

  2. Solid implementation and evidence for the stated operating point. The routing granularity is aligned with tiled GPU execution, and the reported denoising measurements include concurrent geometry work on the same GPU. The dense and model-only sparse comparisons use the same generator and 158-case evaluation protocol; the latter is particularly useful evidence that geometry adds value beyond sparsification alone. The reported quality remains close to dense attention while late-rollout denoising becomes substantially faster. This practical benefit supports weak acceptance even though the evaluation scope is limited.

Critical weaknesses
  1. The evidence has limited coverage and does not establish a universal quality-preserving rule. Evaluation uses one generator, one navigation split, and one routing operating point. Aggregate quality differences are small and lack paired uncertainty estimates, so the slight Overall-score increase should not be interpreted as a demonstrated quality improvement. Geometry failure cases and per-case degradation also deserve clearer characterization. These limitations reduce the strength and generality of the conclusion, but do not undermine the usefulness of the demonstrated operating point.

  2. The scope of the speedup needs to remain precise. The chunk-48 attention and denoising gains are valuable, but they are not measurements of total request latency or peak-memory reduction. Full-rollout costs include other stages and initialization/drain effects; retained history and model-importance computation also remain relevant to overall scaling. A complete wall-clock accounting would make the practical benefit easier to assess. This is a limitation of the system-level evidence rather than a substantive objection to the measured attention/denoising results.

Questions for authors
  1. Runtime clarification: What is the total wall-clock time for the full 48-chunk rollout, including geometry work, initialization, decoding, and pipeline drain, and what is the peak memory usage? Repeated-run variation would help interpret the practical gain.
  2. Existing-output analysis: Can the authors provide paired per-case quality differences or confidence intervals for the 158 cases, and identify the cases with the largest degradation, especially where geometry is unreliable? This would establish whether the near-dense mean hides a systematic failure mode.
  3. Contribution clarification: Beyond the useful model-only sparse baseline, is there a budget-matched recency or field-of-view routing comparison with the same temporal safeguards? Such a result would clarify how much benefit requires the region-level geometric correspondence.
2 comments
AuthorsOpenReview ↗

Response to Reviewer hZum (1/2)

Thank you for the careful and constructive review. We agree that Velox offers a useful operating point rather than a universal rule, and the revision now frames it this way (changes in blue). This part covers the runtime (W2, Q1) and the routing baselines (Q3), and Part 2 covers uncertainty and failure cases (W1, Q2).

W2: The chunk-48 attention and denoising gains are not measurements of total request latency or peak memory. Full-rollout costs include other stages and initialization or drain effects.

Q1: What is the total wall-clock time for the full 48-chunk rollout, including geometry work, initialization, decoding, and pipeline drain, and what is the peak memory? Repeated-run variation would help.

Response to W2 and Q1. We agree that the chunk-48 numbers measure two parts of the rollout rather than the whole request. The revision states what each number covers (Sec. 4.2, Figure 3, and App. B.4 of the manuscript) and what Velox costs beyond it (Sec. 6). The paper's measurement is on one H200, with the geometry worker on the same GPU:

Chunk 48DenseVeloxSpeedup
Attention call (ms)34.27.84.36×
Denoising of one chunk (s)7.192.962.43×

The denoising time is measured while the geometry worker shares the GPU, so it already includes that slowdown. It does not include model loading, the warm-up of the first chunks, VAE decoding, or the final drain (the wait for the geometry worker to finish its last task). Since VAE decoding runs on a side stream off the denoising path (App. C), the main stages outside Figure 3 of the manuscript are loading, warm-up, and the drain.

Beyond these numbers, Velox has two costs that dense attention does not. First, it loads a reconstruction model, and its first chunks run dense attention until the first geometry state is ready. These one-time costs are amortized as the rollout grows, which is where Velox gains most, as the widening per-chunk gap in Figure 3 of the manuscript shows. Second, Velox is designed to cut compute rather than memory. It keeps every past KV block, which preserves the full history for revisits, and also holds the reconstruction model, so its peak memory is higher than that of dense attention.

Q3: Beyond the model-only sparse baseline, is there a budget-matched recency or field-of-view routing comparison with the same temporal safeguards?

Response to Q3. Thank you for suggesting this ablation experiment. Block-Sparse in Table 1 of the manuscript uses the model's own ranking at the same budget and with the same safeguards. Against it, geometry raises the averages by less than one point (API-free 73.11 → 73.51, Overall 75.01 → 75.60) and spatial consistency by 2.5 points (71.8 → 74.3). It also scores higher on perspective consistency (75.7 → 77.0), a metric where both sparse routes fall well below dense attention (81.7). App. A.5 also shows how fusion moves the selected blocks onto the matching region.

Compared with simpler rules, the benefit of geometry lies in revisits. A recency route ignores where a region was seen before, and a field-of-view route gives every region of a chunk the same past frames (Sec. 1 of the manuscript), whereas Velox assigns each region its own history.

AuthorsOpenReview ↗

Response to Reviewer hZum (2/2)

W1: Evaluation uses one generator, one navigation split, and one routing operating point. Aggregate differences are small and lack paired uncertainty, so the slight Overall increase should not be read as a quality gain. Geometry failure cases and per-case degradation deserve clearer characterization.

Q2: Can the authors provide paired per-case quality differences or confidence intervals for the 158 cases, and identify the cases with the largest degradation, especially where geometry is unreliable?

Response to W1 and Q2. Thank you for this suggestion, which we found very useful. The main finding is that the near-dense mean does not hide a failure tied to camera drift. We ran the paired analysis on all 158 cases, using the per-case scores of the runs behind Table 1 of the manuscript, and the full results are in the new App. A.3 and Table 3 of the manuscript. Below are Velox minus dense, in points, with paired bootstrap 95% intervals (10,000 resamples):

MetricCasesVelox − dense
Mean of 8 automatic metrics (App. A.3)158+0.34 [−0.38, +1.08]
Imaging quality158+1.28 [+0.84, +1.74]
Aesthetic quality158+0.55 [+0.06, +1.05]
Spatial consistency60+1.31 [−0.18, +2.83]
Motion smoothness158−0.23 [−0.35, −0.11]
Temporal flickering158−0.28 [−0.45, −0.10]
Subject consistency103−0.81 [−1.58, −0.07]
Perspective consistency103−4.70 [−8.58, −0.94]

Overall and per-metric. The eight-metric mean and 11 of the 17 metrics in Table 3 of the manuscript show no clear difference, and Velox is slightly better on imaging and aesthetic quality. Among the four losses, perspective consistency is the largest (−4.7), while the other three are below one point. Block-Sparse is lower still on this metric in Table 1 of the manuscript (75.7 versus 77.0 for Velox, unpaired), which is consistent with a cost of sparse routing that geometry partly recovers. The drift analysis below also shows that this loss does not grow with camera error.

Worst cases. The ten cases with the largest drop lose 3.8 to 17.2 points on the eight-metric mean. They are spread across the drift groups below (four high, three middle, three low) and do not point to a single failure mode tied to camera error. Interestingly, one of them is a case where dense attention visibly drifts in our qualitative figure, so automatic metrics and visual inspection do not always agree.

Where geometry is unreliable (new App. A.4 and Table 4 of the manuscript). We grouped the 158 cases by how well the generated camera follows the command, using the WBench navigation score (lower means more drift), and fixed the groups before comparing the models. If wrong matches hurt Velox, the high-drift group should show the largest loss:

Drift groupCases8-metric mean
High52+0.41 [−0.94, +1.81]
Middle52+0.42 [−1.01, +1.93]
Low54+0.18 [−0.74, +1.15]

The eight-metric mean shows no clear loss in any group, and perspective consistency is not worse in the high-drift group either (Table 4 in the manuscript). The only clear navigation-score loss is in the low-drift group (−1.68 [−3.18, −0.07], Table 4 in the manuscript). In App. A.4 we also rotate the cameras used by the geometry on purpose. In two of three shortened rollouts, fewer regions pass the coverage test, and those regions fall back to the model-only route.

One generator, one split, one operating point. We have scoped the claims of the revision to this setting.

We hope these results answer your questions, and we are happy to clarify anything further.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.