- 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
-
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.
-
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
-
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.
-
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
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 group | Cases | Mean of 8 automatic metrics |
|---|---|---|
| High | 52 | +0.41 [−0.94, +1.81] |
| Middle | 52 | +0.42 [−1.01, +1.93] |
| Low | 54 | +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:
| Case | No offset | 15° offset |
|---|---|---|
| A (returns to start) | 0.91 | 0.60 |
| B | 0.39 | 0.29 |
| C (bedroom, returns to start) | 0.83 | 0.88 |
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.
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.
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:
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.