{"id":"53fdbe43-9c4b-48f6-9a88-c376c32be067","arxiv_id":"2608.10519","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"SparSTAR dynamically picks which key blocks each attention head reads at expensive video scales, yielding about a 1.6x end-to-end speedup while VBench scores stay within 0.1 points of dense InfinityStar.","lead":"A new training-free sparse-attention method speeds up video generation from autoregressive models while keeping quality nearly unchanged. It recomputes which parts of the video context matter at each scale and clip instead of reusing a fixed sparse pattern.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The fixed density schedule and block-size choice are tuned on a small subset with no held-out sensitivity check, so the 'close-to-dense' and 1.6x claims rest on one operating point; full-suite VBench mitigates but does not fully settle it.","rationale":"I agree with the reader that the block-level aggregated score in Eq. (2) is a key assumption, and it is well supported empirically by Table 11 and Figure 5. The larger unresolved issue is the fixed operating point chosen by ablation on a small dimension-balanced subset, with no held-out sensitivity analysis. This is directly flagged by the paper's conclusion listing 'sensitivity to block size and density schedules' as future work. The full-suite VBench evaluation is impressive and reduces the risk, but it does not make the schedule selection independent of the evaluation. The central claim remains plausible and internally consistent; the weakness is about generalization of the tuned schedule rather than an internal contradiction. Since the reader's verdict is already CONDITIONAL and this concern does not falsify the claim, I recommend no change to the verdict. A held-out prompt evaluation is a concrete, inexpensive way to either confirm or dismiss the concern.","tokens_in":19951,"tokens_out":19255,"duration_ms":211643,"concrete_test":"Run a genuinely held-out evaluation: take the 946 VBench prompts, remove the 96 dimension-balanced prompts used for ablations and the 32 480p analysis prompts, then generate dense InfinityStar and SparSTAR for at least 100 randomly sampled held-out prompts at 720p T2V with the five default seeds. Compute VBench and paired PSNR/SSIM/LPIPS. If the mean VBench gap grows by more than 0.5 points or PSNR drops by more than 1 dB relative to Table 1, the fixed schedule is overfit to the tuning subset; if the gaps stay at headline size, the concern is settled.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that on 720p T2V/I2V, SparSTAR keeps VBench and paired-output fidelity close to dense InfinityStar at about 1.6x speedup. The most load-bearing unverified point is the fixed operating point: block size b=128, sparsified scales s26–s29, and densities (0.625, 0.1875, 0.1875, 0.09375). This schedule is selected by ablation on a dimension-balanced 96-prompt subset (Figure 7, Table 11, and supplementary A.9), and the paper's own conclusion explicitly defers 'sensitivity to block size and density schedules' to future work. Because the final scale retains only 9.375% of its selectable current-scale blocks, a modest shift in attention concentration across prompts could move the VBench/PSNR deltas away from the reported 0.04-point and 31.22 dB values. The full-suite VBench result is broad evidence that the chosen schedule works on the benchmark, but it is not a held-out test of the schedule, and Table 1 reports no error bars. This does not falsify the central claim; it makes its generality conditional on one further check.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SparSTAR, a training-free block-sparse attention method for the InfinityStar spacetime autoregressive video model. SparSTAR partitions queries and keys into 128-token blocks at the expensive late refinement scales, computes per-head aggregated query–key scores S_ij = tau * qbar_i^T kbar_j from current activations, keeps text and SSA reference context dense, and ranks the remaining current-scale and preceding-clip blocks under a fixed per-scale density schedule. It executes the selected blocks through a forward-only FlexAttention path (FastFlex), preserving all tokens and refinement scales. The paper first presents attention-map analyses showing that cross-scale and cross-clip pattern reuse loses substantial attention mass relative to fresh selection, then reports 720p T2V/I2V experiments with about 1.60x/1.62x end-to-end speedup, paired-output PSNR 31.22/27.96 dB, and VBench totals within 0.04/0.08 points of dense InfinityStar, plus 480p and 10-second multi-clip results.","tokens_in":20191,"tokens_out":14044,"duration_ms":124277,"significance":"If the reported results hold, SparSTAR provides a practical training-free acceleration for autoregressive video generation, and its per-scale recomputation of block support is a sensible design choice for the STAR setting. The paper has notable strengths: the attention-mass analyses in Figures 3–5 are clearly described and internally consistent; the evaluation separates the full VBench suite from the dimension-balanced analysis and ablation subsets (Supplementary A.9); dense and sparse runs use matched prompts, seeds, and hardware; and the provenance of the quoted baseline numbers is explicitly disclosed. The main residual risk is that the fixed operating point—density schedule and block size—is selected on a small internal subset and under an implementation constraint, with no held-out sensitivity check, so the generality of the headline fidelity and speedup numbers is conditional on that schedule.","major_comments":[{"comment":"The headline fidelity and speedup claims rest on a single fixed operating point, and neither component is tested in a held-out fashion. The density schedule (0.625, 0.1875, 0.1875, 0.09375) is selected by ablation on the 96-prompt dimension-balanced subset (Figure 7, Table 11), and block size 128 is adopted in Table 12 in part because FlexAttention provides no valid kernel configuration below 128 in your environment. As the paper's own conclusion notes, sensitivity to block size and density schedules is deferred to future work. Because the full VBench suite used for Table 1 overlaps with the subset used for the ablations, the full-suite result is not a clean held-out test of the schedule. Please add a sensitivity analysis on a disjoint prompt split (e.g., perturb the per-scale densities and the block size while reporting VBench/PSNR and latency), and provide error bars or confidence intervals for the headline Table 1 means rather than point estimates only.","section":"Figure 7, Table 12, Supplementary A.9"},{"comment":"The setup text states that VBench evaluates 5-second, 81-frame videos over all 16 dimensions, but Table 3 reports VBench scores for the 480p, 10-second, 161-frame long-video setting. The manuscript does not state whether the 10-second videos are scored directly with the VBench 5-second protocol, truncated to the first 5 seconds, or otherwise adapted. Please specify the exact procedure used for Table 3 and state whether those scores are directly comparable to the 5-second VBench numbers in Tables 1–2 and the supplementary tables.","section":"Quality and efficiency metrics; Table 3"},{"comment":"The baseline rows in Table 1 (SparseVAR, FastVAR, ToMe, FastSTAR) are quoted from the FastSTAR paper, which used a single H100 80GB GPU and a different reconstruction-metric sample (10 videos per dimension) than the full-suite protocol used for SparSTAR. The main-text sentence that SparSTAR 'improves PSNR by 2.92 dB on T2V and 2.31 dB on I2V' presents these as direct deltas, even though neither hardware nor evaluation sample is matched. The disclosure in Supplementary A.8 is helpful, but the main text should either re-evaluate the baselines under the same protocol or rephrase the comparison as being to published numbers from a different setup; otherwise the reader cannot infer a matched relative advantage.","section":"Table 1 and Supplementary A.8"}],"minor_comments":[{"comment":"There is a typo in the Method section: 'Based on the InfinitySTAR' should read 'Based on the InfinityStar'.","section":"Method"},{"comment":"The text says selector arithmetic is 0.016% of dense query-key arithmetic across the sparsified late scales; a naive count with b=128 gives 1/b^2 approximately 0.0061% for the block-pair dot products, so please specify which operations (aggregation, Top-K, mask construction) are included in the 0.016% figure.","section":"Eq. (2) and selector overhead"},{"comment":"Please verify the VBench total formula: the 4:1 weighting of Quality over Semantic in Eq. (8) is unusual relative to the original VBench aggregate script's equal-weight combination of the two perspective scores, and the manuscript should either cite the exact script behavior or correct the formula.","section":"Eq. (8)"},{"comment":"The block-size ablation is partly confounded by the use of a pure-tensor fallback for b=32 and b=64 because no FlexAttention kernel is available; the latency comparison across block sizes is therefore not purely an evaluation of block size, and this limitation should be stated in the main text when the deployed block size is selected.","section":"Table 12"},{"comment":"The nomenclature 'Block-wise Top-K' in Figure 3 and 'Fresh aggregated-QK' in Table 11 should be explicitly defined as the same selection mechanism, since a reader may not immediately connect the two labels.","section":"Figure 3 and Table 11"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the proposed method is interesting. The main gap is the absence of a held-out sensitivity analysis for the fixed density schedule and block size; I would make that a condition of acceptance. I also recommend that the editor ask the authors to clarify the VBench protocol used for the 10-second evaluations, as the reported numbers otherwise look inconsistent with the stated 5-second VBench setup."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The central claim is credible: recomputing a per-scale, per-head block-sparse mask from current query/key activations beats reusing a transferred mask in InfinityStar, and the measured 1.6x end-to-end speedup with VBench deltas under 0.1 points is well supported. This is the first method I have seen that treats sparse attention as a per-scale recomputation problem in STAR video generation, and the attention-mass analyses across scales and clip boundaries are genuinely informative. The forward-only FlexAttention path is a practical detail that makes the dynamic selection affordable, and the paper is refreshingly transparent about what it measured versus what it transcribed: Table 5 explicitly marks FastSTAR's numbers as contextual and not hardware-matched, and the full-suite VBench protocol is clearly documented.\n\nThe analysis sections are the strongest part. The decomposition of quality loss into blockization, aggregated-QK scoring, and reuse-induced mismatch is a clean way to show why fresh selection matters, and the clip-boundary persistence numbers (85% vs 56% retained mass at 15% budget) make the case concretely. Reconstruction metrics are computed on all 4,730 clips for the 720p T2V comparison, which is not the usual handful of examples.\n\nThe soft spots are real but modest. The fixed density schedule (0.625, 0.1875, 0.1875, 0.09375) and block size b=128 are chosen by ablation on a dimension-balanced 96-prompt subset, with no held-out sensitivity check. The paper's own conclusion defers that to future work, and the full-suite VBench is broad evidence that the chosen operating point works on the benchmark, but it is not evidence that the same schedule is near-optimal over varied prompt distributions. A small shift in attention concentration could move the PSNR/VBench deltas more than the headline numbers suggest. That is a conditional-accept concern, not a fatal one. Also, the block score in Eq. (2) is heuristic; empirical validation is good, but there is no bound on how well aggregated-QK ranking tracks token-level attention. Minor. And the code is promised but not yet released, which matters for a method whose main virtue is practical speedup on a specific model.\n\nFor a video generation audience, this is a useful contribution: it identifies a design principle (recompute rather than reuse sparse patterns in autoregressive video) and backs it with careful measurement. It deserves a serious referee. My recommendation: send it out, and ask the authors to release code and add a sensitivity study of the density schedule and block size across held-out prompt subsets. The core result is not in doubt; the generality of the fixed operating point is.","headline":"Solid, honestly-measured sparse-attention method for InfinityStar; the 1.6x/close-to-dense claim holds, but the fixed density schedule deserves a held-out sensitivity check before generalizing.","tokens_in":20735,"tokens_out":1474,"would_cite":true,"duration_ms":15942,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Recomputing sparse attention block selection at every scale lets InfinityStar video generation run about 1.6× faster while keeping VBench quality nearly unchanged.","keywords":["sparse attention","video generation","visual autoregressive modeling","training-free acceleration","block-sparse attention","spacetime autoregressive model","InfinityStar","FlexAttention"],"falsifier":"Run SparSTAR on held-out prompts outside the 96-prompt balanced subset at the fixed density schedule: if the PSNR gap to dense InfinityStar widens substantially beyond the reported level or any VBench dimension drops by more than about one point, the fixed-schedule and block-averaging assumptions are not generally valid. A sharper test replaces the aggregated block score with the exact token-wise Top-K oracle at the same densities: if the oracle's quality advantage over SparSTAR exceeds the paper's measured reuse-mismatch term, then block aggregation, not stale masks, is the dominant source of fidelity loss.","tokens_in":19728,"feed_emoji":"🎬","tokens_out":11928,"duration_ms":93963,"temperature":0.7,"pith_summary":"Speedups from sparse attention in video generation usually come from reusing one sparse pattern across many denoising steps or scales. This paper argues that in InfinityStar's spacetime autoregressive setting, reuse is the wrong move: important key blocks shift from scale to scale and from clip to clip, so the pattern should be recomputed from current query and key activations at each expensive scale and attention head. It presents SparSTAR, a training-free block-sparse attention method that does exactly this, reporting that on 720p text-to-video and image-to-video generation it preserves every token and every refinement scale while delivering about 1.6× end-to-end speedup, with VBench scores within 0.04 and 0.08 points of the dense model. If the claim holds, dynamic block selection is a practical way to accelerate long video autoregressive generation without pruning tokens, skipping scales, or retraining.","feed_headline":"Sparse masks recomputed at every scale give 1.6× video speedup","feed_subtitle":"720p text- and image-to-video quality stays within 0.1 VBench points of the dense model.","key_machinery":"The load-bearing object is the per-head aggregated query-key block score $S_{ij} = \\tau \\bar{q}_i^{\\top} \\bar{k}_j$, where $\\bar{q}_i$ and $\\bar{k}_j$ are block-averaged query and key vectors over contiguous 128-token blocks in attention-kernel order and $\\tau = D^{-1/2}$. Because the dot product of block averages equals the average pairwise pre-softmax compatibility between tokens in the two blocks, this parameter-free score ranks which key blocks matter at each scale and head, and it is computed fresh from current activations rather than transferred. Around the score SparSTAR builds a clip-aware policy: text blocks (and, for Clip 1, the final image-pyramid reference) are always retained dense, while preceding-clip reference and current-scale blocks are jointly ranked under the density schedule $(0.625, 0.1875, 0.1875, 0.09375)$ over the four sparsified refinement scales, with the most aggressive budget on the largest 72,000-token scale. The chosen blocks are executed through a forward-only FlexAttention path, called FastFlex, which writes only the forward mask fields and skips reverse-index construction, keeping selection and mask overhead below 1.7% of end-to-end latency.","core_discovery":"The paper's central discovery is that attention patterns in InfinityStar do not transfer reliably across scales or across clip boundaries, and that the attention-mass loss from reusing them is worse than the latency cost of recomputing them. Reusing a block-selection pattern across scales drops retained attention mass by about 35.6% on average, while recomputing block-wise Top-K at every scale stays within 4.64% of the token-wise upper bound; across a clip boundary at a 15% reference-context budget, fresh selection retains 85% of attention mass versus 56% for reuse. SparSTAR operationalizes this by scoring contiguous 128-token key blocks with a per-head averaged query-key dot product, keeping text (and, for Clip 1, the final image-pyramid reference) dense, jointly ranking preceding-clip and current-scale blocks under a fixed per-scale density schedule, and executing the chosen blocks through a forward-only FlexAttention path. On 720p T2V and I2V this yields 1.60× and 1.62× end-to-end speedups with reconstruction and VBench fidelity close to dense InfinityStar, and the sparse operator is 7.08× faster than FlashAttention-2 while selection and mask construction add less than 1.7% end-to-end latency.","pith_inferences":["As an extension, the same fresh-selection principle should apply to any autoregressive or world-model backbone whose visible key context changes across scales or clips; the paper's evidence suggests mask reuse will decay wherever the reference context is replaced.","The parameter-free block score is a likely bottleneck at very low densities, so a learned or entropy-weighted routing head could improve ranking when the budget is tight; that would trade away the training-free property and is a testable next step.","The fixed density schedule is the least-justified component of the method; per-prompt or per-head dynamic budgets could push the speed-quality frontier further, especially on the final scale where attention mass is most concentrated.","The gap-decomposition metrics the paper introduces could become a standard way to report sparse-attention fidelity, separating the cost of coarse block granularity from the cost of stale masks."],"forward_implications":["If the central claim is correct, the default recipe for sparse attention in video autoregressive models should be to recompute block selection at every scale from current activations instead of transferring masks from earlier scales or from diffusion models.","Because SparSTAR keeps every token and every refinement scale, it composes with branch-level and scale-level accelerations: skipping the unconditional classifier-free-guidance branch on top of SparSTAR reaches about 2.50× speedup on 720p with only a small VBench drop.","Since selection cost scales with the number of 128-token block pairs rather than token pairs, per-scale dynamic recalibration stays practical even at the 72,000-token final scale, where the sparse operator itself runs 7.08× faster than a dense FlashAttention-2 kernel.","The paper's decomposition of the quality gap into blockization, aggregated-QK scoring, and reuse-induced mismatch gives a reusable diagnostic for locating where other sparse attention approximations lose fidelity."],"supporting_citations":[{"why":"Defines the InfinityStar video VAR backbone and its Spacetime Sparse Attention context that SparSTAR sparsifies.","marker":"(Liu et al. 2025)"},{"why":"Provides the FlexAttention programming model through which the forward-only sparse execution path FastFlex runs.","marker":"(Dong et al. 2025)"},{"why":"FlashAttention-2 is the dense attention kernel used as the operator-level baseline for the reported 7.08× sparse speedup.","marker":"(Dao 2024)"},{"why":"Supplies the VBench benchmark used to evaluate video quality in the headline 720p results.","marker":"(Huang et al. 2024)"},{"why":"FastSTAR contributes the paired-output evaluation protocol and the published accelerated-baseline numbers that SparSTAR compares against.","marker":"(Yune, Jeong, and Kim 2026)"},{"why":"SparVAR is the cross-scale mask-transfer method whose reliability assumption the paper's analyses test and reject.","marker":"(Li et al. 2026)"}],"fun_headline_variants":["SparSTAR recomputes attention per scale for 1.6× video speedup","Fresh sparse masks beat reuse in 720p video generation","Scale-wise sparse attention keeps quality at 1.6× speed","Recomputing masks, not reusing, speeds video generation 1.6×","SparSTAR: attention sparsity that adapts to scale for video"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a single score from block-averaged query and key vectors ranks which key blocks matter for generation quality almost as accurately as full token-level attention, and that the same fixed per-scale density schedule works well for every prompt.","fun_headline_variants_meta":{"raw":{"variants":["SparSTAR recomputes attention per scale for 1.6× video speedup","Fresh sparse masks beat reuse in 720p video generation","Scale-wise sparse attention keeps quality at 1.6× speed","Recomputing masks, not reusing, speeds video generation 1.6×","SparSTAR: attention sparsity that adapts to scale for video"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000183,"raw_usage":{"total_tokens":1346,"prompt_tokens":1009,"completion_tokens":337,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":625,"completion_tokens_details":{"reasoning_tokens":236}},"tokens_in":625,"tokens_out":337,"duration_ms":3834,"temperature":1.0,"reasoning_tokens":236,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:18:03.616203+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SparSTAR on held-out prompts outside the 96-prompt balanced subset at the fixed density schedule: if the PSNR gap to dense InfinityStar widens substantially beyond the reported level or any VBench dimension drops by more than about one point, the fixed-schedule and block-averaging assumptions are not generally valid. A sharper test replaces the aggregated block score with the exact token-wise Top-K oracle at the same densities: if the oracle's quality advantage over SparSTAR exceeds the paper's measured reuse-mismatch term, then block aggregation, not stale masks, is the dominant source of fidelity loss.","supporting_citations":[],"review_version":1}