acceptodds
Under review as a conference paper at ICLR 2027

Partition the Support, Reconstruct the Residual: Training-Free Sparse Attention for Video Generation and World Models

Abstract

Block-sparse attention accelerates video transformers by grouping queries and keys and computing attention only between selected groups. However, it faces two challenges: (1) queries sharing a block route may attend to different keys; and (2) retaining most attention mass does not necessarily preserve the attention output. We find that partition choice affects both the overlap among grouped queries’ preferred supports and how well an affine function of the sparse output can represent the difference between dense and sparse outputs. We introduce SparsePR, a training-free method combining Response-Coupled Partitioning with Probe-Fitted Residual Reconstruction. We first group paired key/value tokens by how their keys respond to sampled queries. We then group queries by their responses to the resulting key-group centroids for shared sparse routing. We compute exact attention on a small set of query rows and fit an affine correction to the differences between their dense and sparse outputs using weighted ridge regression. The fitted map predicts a residual for each unprobed query, which we add to its sparse output. We measure attention-output error as the relative L2 error with respect to each query's dense output. Across four video generation and world models, SparsePR reduces this error, averaged over unprobed queries, by 61.6–90.5% compared with semantic block-sparse attention without residual correction, at 22% executed-pair density including exact probes. Ablations show that response-coupled partitioning reduces error before correction and improves reconstruction with the same probe budget, while probe fitting provides the largest overall reduction. SparsePR achieves 1.48×–2.61× end-to-end speedups over dense attention, with dense-reference PSNR of 24.42–31.84 dB at 21.9–26.0% realized executed-pair density.

Primary area:
generative models
Keywords:
sparse attention, efficient inference, video generation, world models, diffusion transformers, training-free acceleration

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_jFULOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

To address the limitations of current block-sparse attention mechanisms, this work introduces the SparsePR framework, which consists of two main components: Response-Coupled Partitioning and Probe-Fitted Residual Reconstruction. Experiments conducted across four different video generation models demonstrate the effectiveness of the proposed approach.

Critical strengths
  1. This method was tested on four video generation models, and the results are quite convincing.
  2. This work includes a comprehensive set of ablation studies, effectively avoiding the pitfall of conflating the model's fit on the training data with its generalization performance.
Critical weaknesses
  1. The presentation of the core end-to-end acceleration results lacks clarity. Could you clarify whether the time measurements for these acceleration results include the time spent on initial clustering, initial eigendecomposition, and weighted ridge regression?
  2. The independent value of the key new contributions, relative to the most relevant existing methods, has not been fully articulated. Given that this approach is plug-and-play and requires no training, I suggest that the authors conduct additional ablation studies by integrating their module into existing methods to demonstrate the performance gains achieved by their proposed approach.
Questions for authors
  1. Clarify which parts the acceleration results include and report the time consumed by each part.
  2. Conduct additional ablation studies by incorporating your proposed module into existing works to demonstrate the performance gains achieved by your method.
3 comments
AuthorsOpenReview ↗

Response to reviewer jFUL

We thank the reviewer jFUL for their comments. We appreciate the reviewer's recognition of our results, evaluation, and generalization performance. We address all concerns below:

W1, Q1. End-to-end latency and component breakdown

The Dense and SparsePR totals in Fig. 4(b) (p. 9) are directly measured end-to-end wall-clock times for Wan2.2. The measured times are 1,650 s and 917 s, respectively. We run both methods in the same H100 allocation, with one untimed warm-up generation followed by one synchronized timed generation per method. In a separate run, we profile the components using CUDA events and apply their shares to the measured wall-clock total. The probe-repair time includes exact probe attention, weighted ridge fitting, and application of the fitted correction. The speedups in Table 1 are based on end-to-end generation times relative to dense attention.

Fig. 4(b) (p. 9) breakdown:

SparsePR componentShare of totalApprox. time (s)
Partitioning and block selection10.8%99.0
Permutation and preparation4.5%41.3
Sparse attention62.4%572.2
Probe repair1.1%10.1
Other stages, including FFNs21.2%194.4
Measured total100%917

Table R8 further provides a complementary comparison of per-head attention latency across correction methods with the routed blocks held fixed

Table R8. Median latency for one full-sequence attention head (ms).

ReconstructionHunyuanVideoWan2.2Cosmos-Predict2.5Cosmos3-Nano
Hard drop5.163.494.191.00
SVG-EAR centroid compensation8.226.036.981.35
SparsePR probe-fitted repair5.794.174.781.30
Centroid compensation + probe fit8.556.427.411.52

W2, Q2. Additional ablation studies

To test partitioning and repair of SparsePR in an existing method, we start from SVG2 [2]. We (1) replace SVG2’s semantic clustering with response-coupled partitioning, (2) add probe-fitted residual reconstruction to the original SVG2, and (3) apply both changes together under SVG2 selection rule. All variants retain SVG2’s top-(p) block selection rule, and we compare them at matched density. Tables R9 and R10 report both attention output error and generation fidelity (PSNR and LPIPS) for HunyuanVideo-13B and Cosmos-Predict2.5-14B, respectively:

Table R9. SparsePR components in SVG2 on HunyuanVideo-13B.

VariantMean rel. L2 ↓Mean per-head p99 ↓PSNR ↑LPIPS ↓
Dense0.000.00--
SVG20.0290.38427.100.114
SVG2 with response-coupled partitioning0.0200.10128.950.104
SVG2 with probe repair0.0200.17929.130.094
SVG2 with both components0.0180.09830.980.085

Table R10 . SparsePR components in SVG2 on Cosmos-Predict2.5-14B.

VariantMean rel. L2 ↓Mean per-head p99 ↓PSNR ↑LPIPS ↓
Dense0.000.00--
SVG20.1800.30818.260.348
SVG2 with response-coupled partitioning0.1630.28323.93 (+5.67)0.227
SVG2 with probe repair0.0560.14723.67(+5.41)0.202
SVG2 with both components0.0350.13525.51 (+7.25)0.081

More ablations related to this question:

  • Matched-density baselines (Tables R1–R2). Compare SparsePR with SVG-EAR[1] and Sol-Attn[3] at matched density.
  • Component ablation (Tables R3–R4). Tables R3 and R4 ablate how partitioning and probe repair affect video fidelity in end-to-end generation.
  • Ablation Comparison with SVG-EAR’s correction (Tables R7–R8). With the partition and routed blocks held fixed, probe repair yields lower mean and p99 attention-output error and lower median per-head latency than SVG-EAR’s centroid compensation on all four models.

[1] Zhou, Xuanyi, et al. "SVG-EAR: Parameter-Free Linear Compensation for Sparse Video Generation via Error-aware Routing." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2026.
[2] Yang, Shuo, et al. "Sparse videogen2: Accelerate video generation with sparse attention via semantic-aware permutation." Advances in Neural Information Processing Systems 38 (2026): 96965-96991.
[3] Li, Haopeng, et al. "Sol-attn: Accelerating video generation inference via on-the-fly attention sparsification." arXiv preprint arXiv:2607.24027 (2026).

I thank the authors for their response. Their explanations have addressed my concerns. I therefore maintain my recommendation to accept the paper. I hope the authors will retain these revisions in the final version.

Dear Reviewer jFUL,

Thank you for taking the time to read our rebuttal and follow up. We’re glad our response addressed your concerns, and we’ll retain these additional experiments and ablations in the final version.

Reviewer_tyxMOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies two limitations of training-free block-sparse attention. Queries that share a route may prefer different keys, and preserving most of the attention mass does not necessarily preserve the attention output. The proposed SparsePR combines response-coupled token partitioning with an affine residual correction. Exact attention is computed for a small set of probe queries, and a weighted ridge model fitted within each attention call predicts the residual between dense and sparse outputs for the remaining queries. Experiments on four video generation and world models report lower attention output error and end-to-end speedups of 1.48x to 2.61x at approximately 22% to 26% executed-pair density.

Critical strengths
  1. The evaluation demonstrates practical gains across several models. Table 1 reports fidelity to dense outputs, generation quality, executed-pair density including probe cost, and end-to-end speedups on four video generation and world models. It also includes comparisons with relevant methods such as SVG-EAR. On every model, SparsePR achieves the highest reported speedup among the methods in Table 1 while keeping generation quality close to dense attention.

  2. The paper includes informative component studies and reports its limitations clearly. Table 2 separates the effects of partitioning and residual correction, and Table 4 shows that the full weighted ridge estimator outperforms mean-only, scalar, and diagonal alternatives. The authors distinguish oracle analysis from prediction and report how often correction increases attention error. These analyses help explain both the method and its limitations.

Critical weaknesses
  1. The contribution of the correction relative to existing compensation methods is not isolated. Residual correction is a central methodological contribution of the paper, and prior sparse attention methods for video generation, such as SVG-EAR, already compensate for skipped interactions. Table 1 compares complete systems, while Table 4 compares the affine estimator only with simpler variants of itself. SparsePR adopts SVG-EAR's routing score while changing both the partitioning and the correction mechanism. Neither comparison therefore shows how much the proposed correction improves over an existing compensation mechanism when the other components are held fixed. Without this comparison, it is difficult to tell whether the gains come from a better correction method or mainly from a different overall configuration.

  2. The generation-level benefit of response-coupled partitioning is not shown. Table 2 establishes improvements in attention output error, but does not show the corresponding changes in generation quality and latency relative to semantic partitioning with the same correction. This comparison would clarify the practical contribution of response-coupled partitioning.

Questions for authors
  1. How does the proposed correction compare with SVG-EAR compensation under matched partitions and selected exact interactions? Where technically compatible, please report reconstruction accuracy and latency with these components held fixed, including the costs of compensation, exact probes, and fitting as applicable.

  2. With the same correction enabled, how do semantic and response-coupled partitioning compare in generation quality and end-to-end latency? Please report this comparison on at least one model.

2 comments
AuthorsOpenReview ↗

Response to reviewer tyxM

W1, Q1 · Reconstruction accuracy and latency with fixed routing

The correction in SVG-EAR[1] and SparsePR differs in what it estimates. SVG-EAR uses key and value centroids to approximate the omitted blocks’ contributions to the softmax denominator and weighted value sum. SparsePR uses exact attention on a small set of probe rows per head to fit an affine correction to the difference between dense and hard-drop outputs. To isolate the effect of the correction, we compare both approaches under the same partition and exact routed blocks held fixed. We further evaluate a hybrid version that combines centroid compensation with probe fitting(this variant is just an exploration). Table R7 summarizes results:

Table R7. Mean / p99 relative L2 attention-output error on the same held-out query rows.

ReconstructionHunyuanVideoWan2.2Cosmos-Predict2.5Cosmos3-Nano
Hard drop0.073 / 0.6960.108 / 0.6470.761 / 7.2270.331 / 3.150
SVG-EAR centroid compensation0.043 / 0.2990.065 / 0.3650.089 / 0.7370.085 / 0.589
SparsePR probe-fitted repair0.030 / 0.2070.063 / 0.3230.075 / 0.4890.076 / 0.447
Centroid compensation + probe fit0.019 / 0.1320.055 / 0.3440.068 / 0.4700.072 / 0.425

Table R8 below reports the median time per attention head for sparse attention and any correction steps used by the method, including centroid compensation for SVG-EAR[1] and exact probe attention and ridge fitting for SparsePR.

Table R8. Median latency for one full-sequence attention head (ms).

ReconstructionHunyuanVideoWan2.2Cosmos-Predict2.5Cosmos3-Nano
Hard drop5.163.494.191.00
SVG-EAR centroid compensation8.226.036.981.35
SparsePR probe-fitted repair5.794.174.781.30
Centroid compensation + probe fit8.556.427.411.52

With the routed blocks held fixed, SparsePR reduces both mean and p99 attention-output error relative to SVG-EAR[1] centroid compensation on all four models, with approximately 4–32% lower median latency per attention head. The hybrid achieves the lowest mean error across all four models and the lowest p99 across three, though this will cause higher latency and is only used as a test.

W2, Q2 .Generation fidelity and partitioning latency

Tables R3 and R4 extend the partitioning ablation to video generation. We compare semantic and response-coupled partitioning with and without probe repair and report PSNR, SSIM, and LPIPS against the dense-attention reference. With the same probe-repair configuration and matched target density, switching to response-coupled partitioning increases PSNR by 1.11 dB and reduces LPIPS by 17.5% on HunyuanVideo-13B-T2V (Table R3). On Cosmos3-Nano, PSNR increases by 1.62 dB and LPIPS falls by 15.6% (Table R4). These fidelity gains follow the same trend as the attention-output error reductions in Table 2.

Table R3. HunyuanVideo-13B-T2V at 22% target density.

VariantPSNR↑SSIM ↑LPIPS ↓Density (%)
Semantic partition + hard drop22.490.7330.24322.00%
Response-coupled partition + hard drop23.310.7630.20322.00%
Semantic partition + probe repair26.840.8790.09722.00%
Response-coupled partition + probe repair (SparsePR)27.950.8800.08022.00%

Table R4. Cosmos3-Nano at 26% target density.

VariantPSNRSSIM ↑LPIPS↓Density (%)
Semantic partition + hard drop21.690.6740.31626.00%
Response-coupled partition + hard drop21.900.6770.28326.03%
Semantic partition + probe repair22.350.7160.22426.00%
Response-coupled partition + probe repair (SparsePR)23.970.7350.18926.03%

We time the partitioning step in isolation. Relative to semantic clustering, response-coupled partitioning reduces partitioning latency by 32.8% on HunyuanVideo, 25.8% on Wan2.2-I2V-A14B, and 29.0% on Cosmos-Predict2.5. On Cosmos3-Nano, partitioning latency increases by 23.1%.

Fig. 4(b) (p. 9) measures end-to-end wall-clock latencies. We run methods in the same H100 allocation, with one untimed warm-up generation followed by one synchronized timed generation per method. For component decomposition, in a separate run, we use CUDA events to profile partitioning and routing, permutation, sparse attention, probe repair, and the remaining non-attention computation. We use the profiled shares to break down the measured wall-clock totals. For full generation on Wan2.2, Fig. 4(b) reports 917 s for SparsePR and 1,650 s for dense attention.

[1] Zhou, Xuanyi, et al. "SVG-EAR: Parameter-Free Linear Compensation for Sparse Video Generation via Error-aware Routing." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2026.

Dear Reviewer tyxM,

Given that the discussion period ends in a few days, we'd appreciate any reviewer feedback on our analysis and experiments provided in the rebuttal, and know if there are any outstanding concerns. Thanks!

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

This paper proposes SparsePR, a training-free block-sparse-attention method for video generation and world models. It groups K/V tokens according to their responses to sampled queries, then groups queries according to responses to K/V-group centroids. SparsePR routes selected blocks sparsely and uses exact attention on (M=64) probe rows per head to fit a weighted ridge-regression predictor of the dense-minus-sparse output residual. The predicted residual is added to non-probe sparse outputs. Experiments on HunyuanVideo, Wan2.2, Cosmos-Predict2.5, and Cosmos3-Nano report dense-reference fidelity improvements at about 22–26% executed-pair density and 1.48–2.61x end-to-end speedups.

Critical strengths
  1. The residual identity in Eq. (3) explains why retained attention mass alone does not determine output error, and the paper explicitly accounts for the exact-probe cost in its executed-pair density. The response-coupled partition and online residual correction address concrete limitations of shared-route sparse attention.
  2. The paper evaluates four large video/world models, reports end-to-end efficiency and output-quality metrics, and includes attention-level ablations for partitioning, probe selection, correction family, and density. These results make the approach promising in practice.
Critical weaknesses
  1. The quality-efficiency advantage is not obtained with fair comparisons. Table 1 compares methods at materially different realized densities, while the density sweep is only against internal variants.
  2. The related-work section cites closely related routing/correction methods such as Sol-Attn, but Table 1 does not compare against them or explain the incompatibility.
  3. The main component ablations are attention-output-error measurements, but not end-to-end generation ablations, so the contribution of each component to final video quality remains unclear.
  4. The reliability of probe-fitted correction on unobserved query groups is not fully studied. SparsePR requests (C_q=300) query groups but uses only (M=64) probes. When many groups are nonempty, Algorithm A.1 leaves many groups without a probe and applies the same global affine correction to them. The paper does not report group/token coverage, error by represented versus unrepresented group, or worst-group behavior.
Questions for authors
  1. With (C_q=300) and (M=64), how many nonempty query groups are there, and what fraction of query tokens are probe-covered for each model? Please report accuracy separately for covered and uncovered groups, as well as worst-group or tail behavior and any corresponding end-to-end generation failures.
  2. Could you provide matched-realized-density quality/latency comparisons with compatible recent baselines, including SVG-EAR and Sol-Attn, or explain why any are incompatible?
  3. Could you provide full-generation ablations for semantic hard drop, response-coupled hard drop, semantic partition plus repair, and SparsePR at equal total executed-pair density?
  4. Figure 4 describes “estimated full-generation latency”; is this directly measured wall-clock latency or extrapolated from component timings? Also, could you clarify the number of prompts, samples, seeds, timing repetitions, and uncertainty estimates behind Table 1?
5 comments
AuthorsOpenReview ↗

Response to reviewer duct

We thank the reviewer duct for their comments and for recognizing the technical strengths of our work. We address your concerns below:

W1, W2, Q2 · Matched-density baseline comparisons

In Table 1 (p. 8), we evaluated each baseline at its own operating point to reflect its chosen quality-latency trade-off. This yieldes to different realized densities across methods. To compare quality and latency at similar attention budgets, we evaluate SparsePR, SVG-EAR[1], and Sol-Attn[2] on HunyuanVideo-13B-T2V. We tune each method’s routing hyperparameters to target similar pair densities of 22% (Table R1) and 30% (Table R2), with realized densities within 0.2 percentage points of one another at each operating point. For each method, we run one untimed warm-up generation and then measure end-to-end generation latency. Timing includes routing and correction overhead for each method. Speedups are relative to dense attention.

Table R1. HunyuanVideo-13B-T2V near 22% realized exact-pair density.

MethodDensityPSNR ↑SSIM ↑LPIPS ↓VBench imaging quality ↑VBench subject consistency ↑E2E speedup ↑
Dense100%———0.6630.9651.00×
SVG-EAR22.002%26.120.8590.1020.6600.9652.01×
Sol-Attn22.104%21.190.7250.2530.6590.9631.87×
SparsePR22.003%27.950.8800.0800.6600.9652.60×

Table R2. HunyuanVideo-13B-T2V near 30% realized exact-pair density.

MethodExact-pair densityPSNR ↑SSIM ↑LPIPS ↓VBench imaging quality ↑VBench subject consistency ↑E2E speedup ↑
Dense100%———0.6630.9651.00×
SVG-EAR30.18%29.620.8900.0680.6600.9601.86×
Sol-Attn30.26%24.170.8010.1780.6590.9631.33×
SparsePR30.07%30.860.9190.0550.6600.9662.25×

Relative to SVG-EAR[1], SparsePR improved PSNR by 1.83 dB at 22% density (Table R1) and by 1.24 dB at 30% density (Table R2). At these two operating points, SparsePR achieved end-to-end speedups of 2.60× and 2.25× over dense attention, compared with 2.01× and 1.86× for SVG-EAR[1]. This corresponds to 23% and 17% lower generation latency, respectively. SparsePR also achieved the highest SSIM and lowest LPIPS among the sparse methods at both densities.

W3, Q3 · End-to-end ablation of partitioning and repair

We evaluate semantic hard drop, response-coupled hard drop, semantic partition plus repair, and SparsePR on HunyuanVideo-13B-T2V and Cosmos3-Nano. We target total densities of 22% and 26%, respectively, close to SparsePR’s operating points in Table 1 (p. 8). Within each model, all four variants share the same target density, including exact probe attention where applicable. Across the tested variants on both models, lower attention-output error is accompanied by better video fidelity to the dense-attention reference.

Table R3. HunyuanVideo-13B-T2V at 22% target density.

VariantPSNR↑SSIM ↑LPIPS ↓Density (%)
Semantic partition + hard drop22.490.7330.24322.00%
Response-coupled partition + hard drop23.310.7630.20322.00%
Semantic partition + probe repair26.840.8790.09722.00%
Response-coupled partition + probe repair (SparsePR)27.950.8800.08022.00%

Table R4. Cosmos3-Nano at 26% target density.

VariantPSNRSSIM ↑LPIPS↓Density (%)
Semantic partition + hard drop21.690.6740.31626.00%
Response-coupled partition + hard drop21.900.6770.28326.03%
Semantic partition + probe repair22.350.7160.22426.00%
Response-coupled partition + probe repair (SparsePR)23.970.7350.18926.03%

With probe repair enabled, switching from semantic to response-coupled partitioning raises PSNR by 1.11 dB and lowers LPIPS by 17.5% on HunyuanVideo. On Cosmos3-Nano, it raises PSNR by 1.62 dB and lowers LPIPS by 15.6%. Adding probe repair improves all three fidelity metrics under both partition types on both models. These results extend the ablations from attention outputs to end-to-end video fidelity.

AuthorsOpenReview ↗

Response to reviewer duct (continued)

W4, Q1 · Probe coverage and reliability on uncovered groups

We fit one affine correction per head per attention call and apply it to all unprobed rows, including those in groups without probes (Sec. 4.2). When at least 64 query groups are nonempty, Algorithm A.1 selects one probe from each of the 64 largest groups.

We replay 32 attention calls per model across all query heads at a 22% target density. Token coverage is the fraction of query tokens in covered groups, defined as groups with at least one probe. We report errors on held-out, unprobed rows. Table R5 measures probe coverage and mean attention-output error after repair.

Table R5. Probe coverage and mean attention-output error after repair.

ModelNonempty groups, median [min–max]Tokens in covered groupsCovered mean relative L2 ↓Uncovered mean relative L2 ↓
HunyuanVideo-T2V-13B300 [300–300]31.30%0.021650.03473
Wan2.2-I2V-A14B300 [297–300]31.61%0.047460.07120
Cosmos-Predict2.5-14B300 [40–300]36.23%0.049710.09003
Cosmos3-Nano-16B300 [160–300]34.27%0.063110.08345

To measure the effect of repair in uncovered groups, we further compare hard drop and probe repair on the same held-out rows and routed blocks (Table R6):

Table R6. Relative L2 attention-output error in uncovered groups (hard drop → probe repair)

ModelMean row error ↓p95 of group means ↓
HunyuanVideo-T2V-13B0.06151 → 0.034730.25302 → 0.12746
Wan2.2-I2V-A14B0.11307 → 0.071200.36892 → 0.19450
Cosmos-Predict2.5-14B0.76778 → 0.090033.47464 → 0.27959
Cosmos3-Nano-16B0.43794 → 0.083451.81349 → 0.24019

Across the four models, probe repair reduces mean error on uncovered rows by 37–88% and the 95th percentile of mean error across uncovered groups by 47–92%.

Tail behavior and error increases:

In Appendix D.6, we report that repair increases mean attention error in 1.41–13.98% of attention head evaluations, depending on the model. For example, at layer 20 of Cosmos3-Nano, mean error rises by 2.09% across the evaluated heads and denoising steps.

We also examine every row in the uncovered groups for one Wan2.2-I2V attention call at layer 2, denoising step 6. In this case, repair raises the 95th percentile of mean relative L2 across uncovered groups from 0.0424 to 0.0479 and the maximum from 0.2585 to 0.2616. In the corresponding videos we inspected end-to-end, we did not observe visible degradation in video quality.

Q4. Latency measurement and evaluation protocol.

The Dense and SparsePR totals in Fig. 4(b) (p. 9) are directly measured end-to-end wall-clock latencies. We run both methods in the same H100 allocation, with one untimed warm-up generation followed by one synchronized timed generation per method. For component decomposition, in a separate run, we use CUDA events to profile partitioning and routing, permutation, sparse attention, probe repair, and the remaining non-attention computation. We use the profiled shares to break down the measured wall-clock totals. For evaluation protocol in table 1, we evaluate 50 examples for each model, task, and benchmark configuration (VBench, VBench-I2V, PBench). Each input uses one fixed seed shared across all methods.

[1] Zhou, Xuanyi, et al. "SVG-EAR: Parameter-Free Linear Compensation for Sparse Video Generation via Error-aware Routing." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2026.
[2] Li, Haopeng, et al. "Sol-attn: Accelerating video generation inference via on-the-fly attention sparsification." arXiv preprint arXiv:2607.24027 (2026).

Dear Reviewer duct,

Given that the discussion period ends in a few days, we'd appreciate any reviewer feedback on our analysis and experiments provided in the rebuttal, and know if there are any outstanding concerns. Thanks!

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary and Reviewer Updates

We thank the reviewers for recognizing SparsePR’s practical gains across four video generation and world models, as well as highlighting our ablation studies and observations on sparse attention. In response to the reviewers’ concerns, we added comparisons at matched densities with SVG-EAR and Sol-Attn on HunyuanVideo at two different densities (Tables R1 and R2), full-video generation ablations of partitioning and repair on HunyuanVideo and Cosmos3-Nano (Tables R3 and R4), analyses of probe coverage and repair performance in groups without probes across all four models (Tables R5 and R6), comparisons of accuracy and latency between SparsePR’s probe-fitted repair and SVG-EAR’s centroid compensation using identical partitions and routed blocks across all four models (Tables R7 and R8), and ablations evaluating our partitioning and repair components within SVG2 on HunyuanVideo and Cosmos-Predict2.5 (Tables R9 and R10). These experiments demonstrate the independent benefits of partitioning and repair. We also clarified end-to-end timing, overhead accounting, and component profiling. Following these responses, reviewer tyxM raised their rating from 2 to 3, and reviewers jFUL and duct maintained their ratings of 3, bringing the scores to 3,3,3. Reviewers tyxM and jFUL explicitly stated that their concerns were addressed. Reviewer duct acknowledged that the rebuttal strengthened the paper and maintained their positive recommendation. They raised follow-up questions about comparisons at matched densities across more models, timing reliability, and robustness in groups without probes. Our final reply addresses these points by clarifying the available evidence and its limits. We will incorporate the additional experiments, ablations and discussion into the revised manuscript. We appreciate all reviewers’ time and helpful feedback throughout the rebuttal process.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.