acceptodds
Under review as a conference paper at ICLR 2027

PulseQuant: Propagation-Guided Subspace Correction for 4-Bit Video Diffusion Transformers

Abstract

Quantization errors in video diffusion transformers can be amplified or attenuated by subsequent denoising updates, making local reconstruction error an incomplete predictor of final impact. We introduce PulseQuant, a 4-bit post-training quantization method that combines trajectory sensitivity with activation geometry to guide offline calibration. Isolated block–step interventions estimate propagation risk, which prioritizes sensitive trajectory states during row-radius selection. With these radii fixed, response-subspace correction uses neighboring-code edits to reduce residual components along dominant activation directions. Both stages preserve the original 4-bit weight representation. Controlled interventions show that short-horizon propagated error predicts final latent error more reliably than immediate block-output error, supporting calibration beyond local reconstruction objectives. Evaluations on Wan models, Self Forcing, and MiniMax-H3 demonstrate improvements in key consistency and dense-reference metrics while remaining competitive on other attributes across model scales and generation paradigms.

Primary area:
generative models
Keywords:
Post-Training Quantization, Low-Bit Quantization, Efficient Video Generation

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

This paper proposes a post-training quantization framework for video diffusion transformers, noting that quantization errors accumulate continuously during the denoising process, making it difficult to predict the final generation quality based on local quantization errors. PulseQuant introduces isolated block–timestep pulses during the diffusion process. The propagation gains estimated through quantization-error propagation are incorporated as propagation-risk weights in row-radius calibration. Second-order moments of the activations are estimated using calibration data, and two directions with high activation energy are identified through an iterative process. Neighboring-code edits are then performed along these two activation directions to reduce response error. On the Wan models, Self Forcing, and MiniMax-H3, PulseQuant improves dense reference fidelity metrics (PSNR/SSIM/LPIPS) and overall consistency at the W4A4 level, remains competitive at the W4A6 level, and demonstrates the advantages of native NVFP4 deployment as well as complementary acceleration with low-bit attention.

Critical strengths
  1. Using isolated block–timestep pulses to measure the propagation of quantization error across multiple time steps is an effective method for determining the selection of weight radii. This approach addresses the issue that optimizing for local quantization error does not effectively translate to optimizing for the final quantization error. The two selected directions with high activation energy also effectively guide the editing of the neighboring-code.

  2. Evaluation results across multiple video generators demonstrate that PulseQuant exhibits strong quantitative performance compared to baselines such as OrbitQuant and DVD-Quant. Ablation experiments also validate the effectiveness of each module. Deployment results on an RTX 5080 with native NVFP4 and two attention backends further show reductions in peak memory usage and latency.

Critical weaknesses
  1. When performing Response-Subspace Correction, selecting only two directions with high activation energy may lead to insufficient calibration when activation energy is relatively dispersed. Neighboring-code edits also allow adjustments only within adjacent ranges, which may prevent the calibration from achieving its full potential.

  2. Complex calibration processes—such as the injection of quantized pulses at specified block–timesteps, row-radius searches, the selection of high-energy activation directions, and neighboring-code edits—introduce calibration time overhead. This overhead may increase significantly, particularly as the number of anchor grids and the number of full-precision BF16 denoising steps HH increase.

Questions for authors
  1. If a larger range of index shifts were allowed, or if more active directions were protected, would the final quantization quality improve further? Would this improvement justify the additional calibration cost?

  2. What is the offline calibration cost? Can the authors analyze the calibration time for at least one model? How does the calibration cost vary with the anchor grid and the number of full-precision denoising steps HH?

  3. How sensitive are the results to the choice of the OrbitQuant codebook? Is subspace correction still effective when using a standard uniform quantizer per channel or other codebooks?

3 comments
AuthorsOpenReview ↗

Response to W1&Q1: Additional correction directions and wider code-index edits.

We evaluated both extensions on Wan-1.3B W4A6: more correction directions (k=1,2,4k=1,2,4, with the edit range fixed at ±1\pm1) and wider code-index edits (±1\pm1, ±2\pm2, ±4\pm4, with k=2k=2). All runs share the same calibration prompts, risk profile, radius-selection procedure, and 326 evaluation prompts with matched seeds and BF16 references.

More directions. The differences are small and metric-dependent. Relative to k=2k=2, k=4k=4 improves four of nine metrics (AQ, MS, SC, and PSNR) and slightly worsens the other five (IQ, BC, DD, SSIM, and LPIPS); k=1k=1 has the best LPIPS. k=2k=2 is best on IQ, DD, BC, and SSIM, so we keep it as a balanced default rather than claim it is optimal for every metric.

DirectionsIQ ↑AQ ↑MS ↑DD ↑BC ↑SC ↑PSNR ↑SSIM ↑LPIPS ↓
k=1k=165.5764.7298.3265.2795.4093.9015.600.5330.410
k=2k=265.7864.9098.3766.1195.6693.9215.730.5340.411
k=4k=465.5664.9498.4565.2795.6294.7115.760.5330.412

Wider edits. Expanding from ±1\pm1 to ±2\pm2 improves AQ, BC, SC, PSNR, and SSIM, but worsens IQ, DD, and LPIPS; expanding further to ±4\pm4 gives no consistent additional benefit. The default ±1\pm1 has the best IQ, DD, and LPIPS and ties for the best MS, which supports it as a conservative choice.

Edit rangeIQ ↑AQ ↑MS ↑DD ↑BC ↑SC ↑PSNR ↑SSIM ↑LPIPS ↓
±1\pm165.7864.9098.3766.1195.6693.9215.730.5340.411
±2\pm265.6864.9298.3763.8895.7694.3515.760.5350.415
±4\pm465.3864.7398.3163.8895.7194.4115.760.5300.416

Both extensions act only during offline calibration: they keep valid quantization codes and fixed row radii, and change neither the deployed representation nor the inference operations. The extra cost is small: correction takes 12.05 s for the default k=2,±1k=2,\pm1, 12.21 s for k=4,±1k=4,\pm1, and 13.74 s for k=2,±4k=2,\pm4. However, the gains are at most 0.03 dB PSNR and mixed across metrics, so we keep the default.

AuthorsOpenReview ↗

Response to W2&Q2: Offline calibration cost.

We measured Wan-2.1-T2V-1.3B W4A6 calibration on a single NVIDIA A800 80GB GPU with three calibration prompts. With a 4×44\times4 anchor grid, T=50T=50 denoising steps, and propagation horizon H=4H=4, calibration computation takes 713.41 s (0.198 GPU-hours): 603.53 s for risk profiling, 89.11 s for activation collection, 5.03 s for radius selection, 12.05 s for subspace correction, and 3.68 s for other operations. Including model loading, warm-up, cache writing, and process startup, the total is 747.32 s (0.208 GPU-hours).

The cost scales as follows (total GPU-hours, other settings at their defaults):

Varied settingValuesTotal GPU-hours
Anchor grid (T=50T=50, H=4H=4)4×24\times2, 4×44\times4, 4×84\times80.172, 0.208, 0.280
Denoising steps TT (4×44\times4, H=4H=4)10, 25, 500.171, 0.185, 0.208
Horizon HH (4×44\times4, T=50T=50)1, 2, 4, 80.176, 0.188, 0.208, 0.249

Profiling dominates the cost and grows sublinearly with the anchor grid and with HH in this range, while radius selection and correction are comparatively inexpensive. We will add these measurements to the Appendix.

AuthorsOpenReview ↗

Response to Q3: Subspace correction with a uniform quantizer.

We tested this directly on Wan-1.3B by replacing the OrbitQuant weight codebook with a per-output-channel uniform INT4 quantizer, while keeping the rotation, risk-weighted scale selection, and the A6 activation quantizer.

Uniform INT4IQ ↑AQ ↑MS ↑DD ↑BC ↑SC ↑PSNR ↑SSIM ↑LPIPS ↓
Without correction61.4361.9498.2459.7294.9593.3915.050.5060.479
With correction62.2862.3198.2462.5095.0993.6715.080.5090.472

Correction improves or matches all nine metrics: PSNR rises from 15.05 to 15.08, SSIM from 0.506 to 0.509, and LPIPS improves from 0.479 to 0.472, with larger gains on IQ (+0.85) and DD (+2.78). Subspace correction therefore remains effective beyond the OrbitQuant codebook, under a standard per-channel uniform quantizer.

Reviewer_2Aq9OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

PulseQuant calibrates four-bit video diffusion transformers using trajectory sensitivity and activation geometry. Isolated block–timestep perturbations estimate downstream amplification, which weights row-radius selection. A second stage adjusts neighboring quantization codes along activation-informed response directions. Both stages operate offline and retain the compressed weight representation. Experiments span five video generators, two activation precisions, and component ablations. W4A4 results show consistent fidelity gains over the tested baselines. Native deployment measurements demonstrate memory savings and additional DiT acceleration alongside low-bit attention.

Critical strengths
  1. The method addresses a meaningful limitation of local calibration. It combines trajectory-sensitive radius selection with activation-informed code edits. Both stages operate offline, retaining four-bit weights without an additional high-precision residual branch.

  2. W4A4 fidelity gains extend across five generators. PulseQuant leads the tested baselines on PSNR, SSIM, and LPIPS in every model. Against OrbitQuant, PSNR improves by 0.16–0.45 dB. Matched prompts and seeds strengthen these comparisons.

  3. The ablation supports complementary benefits. Risk weighting raises PSNR from 14.97 to 15.59 dB. Correction alone reaches 15.61 dB, and their combination reaches 16.08 dB. The response-direction controls further support activation-adapted correction.

Critical weaknesses
  1. The quality and deployment paths need a clearer connection. Quality tables use simulated spherical-codebook quantization, while timing uses packed NVFP4. Equivalent reconstructed weights and activations are not established. It is therefore unclear whether the fidelity and speed results describe the same deployed model.

  2. Figure 1 uses different comparison protocols. Local errors are pooled, while propagated errors are ranked within timesteps. Calibration also uses an amplification ratio rather than the plotted propagated error. This leaves the evidence for the specific weighting rule less direct.

  3. Offline cost is difficult to assess. Profiling requires dense continuations at 80 block–timestep anchors, followed by activation collection and code correction. Calibration runtime and peak memory are not reported. These costs affect the practical comparison with simpler calibration methods.

Questions for authors
  1. Figure 1 compares pooled local errors with within-timestep propagated-error ranks. Could you summarize both correlations using identical timestep stratification from the existing interventions? Please also explain how the amplification ratio in Equation 4 tracks final error. This would clarify why the selected weights improve upon local reconstruction.

  2. Tables 1–2 use Q/DQ, whereas Figure 6 times packed NVFP4. Does native export preserve the shared codebook and calibrated row radii exactly? Please clarify activation format and any conversion or recalibration, using existing numerical checks if available. This would establish whether quality and speed results apply to the same model.

3 comments
AuthorsOpenReview ↗

Response to W1&Q2: Connection between quality evaluation and native deployment.

We agree that the two paths need a clearer connection. Tables 1-2 evaluate PulseQuant and all quantized baselines with Q/DQ simulation, which gives a common protocol for their respective quantizers; Figure 6 measures packed-NVFP4 execution speed.

To answer the question directly: native export does not preserve the spherical codebook and calibrated row radii exactly. The native path uses NVFP4 W4A4, and the spherical weights are converted to NVFP4 without recalibration. On Wan-1.3B this conversion introduces a mean per-layer relative weight MSE of 0.00336 across the 300 quantized layers. For the same NVFP4 format, however, native execution matches its Q/DQ reference: the maximum relative output MSE across export tests is 2.17×10−112.17\times10^{-11}.

End-to-end quality at W4A4 on 20 randomly selected prompts:

MethodIQ ↑AQ ↑MS ↑BC ↑SC ↑PSNR ↑SSIM ↑LPIPS ↓
Q/DQ simulation60.3463.3098.6597.5195.5518.370.5190.439
NVFP4 native59.7963.0398.6997.1995.4817.870.5040.457

The native path loses 0.50 dB PSNR, 0.015 SSIM, and 0.018 LPIPS, while the VBench attributes change by at most 0.55 points. Because the radii and code corrections were optimized for the spherical quantizer, they may be suboptimal for packed NVFP4; recalibrating under the native format could recover part of this loss, but we have not yet verified this. We will separate the two settings in the paper, add these measurements, and list native-format recalibration as a limitation.

AuthorsOpenReview ↗

Response to W2&Q1: Matched correlation protocols and the amplification-based weighting rule.

Thank you for identifying the protocol mismatch in Figure 1. Comparing pooled local-error correlations with within-timestep propagated-error correlations does not isolate the benefit of propagation information.

Why amplification. Eq. 6 already evaluates the local reconstruction error of each candidate radius, so the weight is meant to add what that term lacks: how strongly an error at a given state is amplified downstream. EHE_H, in contrast, also reflects the initial magnitude of the profiling perturbation. This motivates gg, but does not by itself show that gg-derived weights outperform EHE_H-derived ones, so we tested both.

Matched protocol. We ran a supplementary analysis with 256 interventions (4 prompts × 2 seeds × 4 timesteps × 8 blocks), separate from the 80 interventions in Figure 1, and report pooled and within-timestep Spearman correlations for every predictor on this same set.

PredictorFinal error: pooledFinal error: within-timestepFinal amplification: pooledFinal amplification: within-timestep
Local reconstruction MSE-0.2160.068-0.301-0.421
Propagated error EHE_H-0.7220.697-0.878-0.168
Amplification gain gg-0.0990.458-0.1440.797
Saved profile weight ω\omega0.5860.3490.6710.664

Here, final amplification is Efinal/E0\sqrt{E_{\mathrm{final}}/E_0}. Under identical within-timestep stratification, local reconstruction MSE is essentially uninformative about final error (0.068), EHE_H is the strongest predictor of absolute final error (0.697), and gg is the strongest predictor of relative amplification (0.797). The sign reversal of EHE_H between pooled and within-timestep correlations is consistent with a mixing effect: differences in error scale and remaining propagation across timesteps dominate the within-timestep association once interventions are pooled, which is why matched stratification matters. The weight actually used in calibration, ω\omega, is the only predictor that correlates positively with both targets under both protocols (0.35-0.67).

End-to-end comparison. We additionally compared gg-based and EHE_H-based weighting on Wan-1.3B W4A6, with all other settings fixed and the same 326 prompts, seeds, and BF16 references.

WeightingPSNR ↑SSIM ↑LPIPS ↓AQ ↑BC ↑SC ↑Scene ↑
gg15.640.5270.41164.7995.4493.1043.97
EHE_H15.500.5140.43264.6295.3693.7244.93

gg gives better dense-reference fidelity on all three metrics, as well as AQ and BC, while EHE_H is better on SC and Scene. These results support gg as a reasonable amplification-based choice; they do not establish universal superiority over EHE_H. We will add both analyses in the Appendix.

AuthorsOpenReview ↗

Response to Q3: Calibration runtime and peak memory.

We agree that the offline cost should be reported. We measured synchronized runtime and peak PyTorch-allocated GPU memory for W4A6 calibration on A800-80GB GPUs. Wan-1.3B and Wan-14B use a single GPU, while MiniMax-H3 uses 8 GPUs.

ModelAnchorsProfiling (s)Radius selection (s)Correction (s)End-to-end (s)GPU-hoursPeak memory (GiB/GPU)
Wan-1.3B4 × 4603.535.0312.05740.090.20619.14
Wan-14B10 × 89,322.2018.0443.769,948.262.76343.80
MiniMax-H310 × 82,787.1123.70105.893,290.297.31275.76

Profiling dominates the cost (82-94% of end-to-end time); radius selection and correction are comparatively cheap. End-to-end time also includes activation collection, model loading, and cache serialization. These are one-time offline costs for a fixed calibration configuration, amortized over all subsequent generation, but they are additional overhead relative to simpler calibration methods. We will add this table in the Appendix.

Reviewer_KDZzOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes PulseQuant, a 4-bit post-training quantization (PTQ) method for video diffusion transformers. It builds on the OrbitQuant representation and adds two offline calibration stages that keep the deployed 4-bit format unchanged. The experiment shows that local block-output RMSE is essentially uncorrelated with final latent error, while the 4-step propagated error recovers the within-timestep ranking of final error.

Critical strengths
  1. A useful diagnostic of the local-to-global mismatch.

  2. Format-preserving method with broad, up-to-date evaluation and a real deployment path.

Critical weaknesses
  1. The contribution of propagation risk is not isolated, and the radius objective cannot use the evidence that motivates it. the paper's central claim is "propagation-guided" calibration. In this paper, most of the gain could come from a generic radius search on top of OrbitQuant.

  2. mixed gains over the strongest baseline, no variance estimates, and a gap between the evaluated and deployed numerics.

Questions for authors

Q1: Please report the radius search with \omega ≡ 1, with a simple timestep-decay weight, with local-error weights, and with the proposed g.

Q2: What part of \omega is actually used. Given the per-layer scale invariance of Eq. 6, how much do the timestep profiles of \omega differ across blocks?

Q3: How many prompts and seeds are used for VBench and for the paired metrics, and which prompt list?

Q4: Table 1 (PulseQuant, W4A6, Wan-1.3B: PSNR 15.64, IQ 64.96) differs from Tables 3/7 (PSNR 16.08, IQ 65.83). What is the "300-layer coverage" protocol, how does it differ from the main tables, and what does OrbitQuant achieve under it?

Q5: Please report γ and [a, b], and how profiling is done for Self Forcing.

Q6: Please report total calibration cost in GPU-hours.

Q7: What SVDQuant rank, ViDiT-Q precision setting, and calibration data were used for the data-dependent baselines? Was the same layer coverage used for every method?

6 comments
AuthorsOpenReview ↗

Response to W1&Q1: Separating radius selection from propagation weighting.

Thank you for raising this distinction. Radius selection and propagation weighting are separate components: the former searches row radii, whereas the latter only changes the relative importance of calibration states inside the same search objective. We agree that the original ablation labels did not separate them clearly. The "+ Rotation" row of Table 3 ("Rotation only" in Table 7) already performs radius selection with uniform weights (ω≡1\omega\equiv1), and "+Risk" changes only the state weights while keeping the same radius search.

To isolate the two effects, we ran the requested variants on Wan-1.3B W4A6 under the ablation setting of Tables 3/7 (λ=0.5\lambda=0.5, ρ=0.25\rho=0.25, no subspace correction). "Rotation, no radius search" keeps the initial row radii; "Radius, ω=1\omega=1" and "Radius, proposed gg" are exactly the "Rotation only" and "Risk only" rows of Table 7.

VariantIQ ↑AQ ↑MS ↑BC ↑OC ↑PSNR ↑SSIM ↑LPIPS ↓
Rotation, no radius search63.5163.4498.0595.0624.6614.910.5090.437
Radius, ω=1\omega=163.2062.3998.0995.0124.7014.970.5120.443
Radius, decay63.8464.9998.1695.2525.4215.340.5210.418
Radius, local error63.9163.6198.2195.1424.6014.740.5090.434
Radius, proposed gg64.3864.9798.1895.5825.7315.590.5290.414

A generic radius search contributes little by itself: with uniform weights it changes PSNR by only +0.06 dB over rotation without radius search (14.91 to 14.97) and is slightly worse on IQ, AQ, BC, and LPIPS. Most of the gain comes from the weighting: the proposed gg adds a further +0.62 dB (14.97 to 15.59) and reduces LPIPS from 0.443 to 0.414. A simple timestep-decay weight recovers part of this gain (15.34 dB), whereas local-error weights do not help (14.74 dB). The proposed gg is best on IQ, BC, OC, PSNR, SSIM, and LPIPS; decay is marginally higher on AQ and local-error weighting on MS. We will relabel the ablation rows and add this table to the paper.

AuthorsOpenReview ↗

Response to W2: Mixed gains over strong baselines & Variance estimates & Quality evaluation versus deployment numerics

W2.1: Mixed gains over strong baselines.

We agree that PulseQuant does not improve every individual VBench attribute. Its advantage is clearest in dense-reference fidelity under the more challenging W4A4 setting: compared with OrbitQuant, the strongest baseline, PulseQuant has better PSNR, SSIM, and LPIPS on all five evaluated models, with PSNR gains of 0.16-0.45 dB. These paired metrics directly measure how faithfully quantization preserves the corresponding full-precision generation. At W4A6, the two methods are closer. We will state this explicitly rather than describe the gains as uniform.

W2.2: Variance estimates.

We address variance at two levels: variation across evaluation prompts, and sensitivity to calibration prompts and seeds.

Evaluation uncertainty. We performed 20,000 paired prompt-level bootstrap resamples of the existing PulseQuant and OrbitQuant results and report 95% percentile confidence intervals for the difference (PulseQuant minus OrbitQuant).

Model / W&AΔ\Delta PSNR (dB) ↑Δ\Delta SSIM ↑Δ\Delta LPIPS ↓
Wan-1.3B W4A6+0.3352 [0.1609, 0.5045]+0.0052 [-0.0029, 0.0133]-0.0132 [-0.0211, -0.0053]
Self Forcing W4A6+0.1524 [0.0251, 0.2802]+0.0202 [0.0150, 0.0255]-0.0112 [-0.0175, -0.0049]
Wan-14B W4A6-0.0290 [-0.2380, 0.1768]-0.0060 [-0.0140, 0.0020]+0.0059 [-0.0030, 0.0149]
Wan-14B W4A4+0.4552 [0.2515, 0.6572]+0.0110 [0.0038, 0.0182]-0.0112 [-0.0196, -0.0030]
Wan-2.2 W4A4+0.3232 [0.1205, 0.5297]+0.0101 [0.0010, 0.0191]-0.0070 [-0.0168, 0.0028]
MiniMax-H3 W4A4+0.3077 [0.1234, 0.4917]+0.0150 [0.0084, 0.0216]-0.0033 [-0.0116, 0.0048]

The point estimates agree with the differences in Tables 1-2. The PSNR interval excludes zero in five of the six settings, the SSIM interval in four, and the LPIPS interval in three. On Wan-14B W4A6 all three intervals contain zero, so the two methods are not distinguishable there. In all three W4A4 settings the PSNR and SSIM gains are significant; the LPIPS gain is significant on Wan-14B and directionally favorable but not significant on Wan-2.2 and MiniMax-H3.

Calibration stability. Separately, we ran nine calibrations on Wan-1.3B W4A6 (three calibration-prompt groups × three calibration seeds) with a 4×44\times4 anchor grid, H=4H=4, and 50 denoising steps.

Calibration variationMean block-risk Spearman correlationRow-radius selection agreement
Prompts only0.928173.48%
Seeds only0.923280.90%
Both0.891971.63%

Block-risk rankings are stable across calibration conditions (Spearman 0.89-0.93), while individual row-radius choices are more sensitive (72-81% agreement).

The two analyses are complementary: the bootstrap intervals quantify evaluation-prompt uncertainty conditional on the recorded models and seeds, whereas the calibration study measures sensitivity to calibration choices.

W2.3: Quality evaluation versus deployment numerics.

We acknowledge this gap. Our quality tables use simulated (Q/DQ) spherical-codebook quantization, whereas the deployment measurements use packed NVFP4, and the two paths are not numerically identical. We have now measured the difference on Wan-1.3B:

  • Exporting the spherical weights to NVFP4 without recalibration introduces a mean per-layer relative weight MSE of 0.00336 across the 300 quantized layers.
  • For the same NVFP4 format, native execution matches its Q/DQ reference: the maximum relative output MSE across export tests is 2.17×10−112.17\times10^{-11}.
  • End-to-end, on 20 randomly selected prompts at W4A4, the native path loses 0.50 dB PSNR relative to the simulated path (table below).
MethodIQ ↑AQ ↑MS ↑BC ↑SC ↑PSNR ↑SSIM ↑LPIPS ↓
Q/DQ simulation60.3463.3098.6597.5195.5518.370.5190.439
NVFP4 native59.7963.0398.6997.1995.4817.870.5040.457

The gap does not affect the comparisons in Tables 1-2: all methods, including OrbitQuant, are evaluated under the same Q/DQ simulation with identical layer coverage, as is common practice in PTQ evaluation, so the relative gains are measured under a single protocol. We will separate the quality and speed evaluation settings explicitly, add these measurements, and list native-format recalibration as a limitation.

AuthorsOpenReview ↗

Response to Q2 & Q3 & Q4

Q2: What information in ω\omega affects radius selection, and how much do its temporal profiles differ across blocks?

The reviewer is correct about the per-layer scale invariance of Eq. 6. Our implementation makes this explicit by normalizing the weights to unit mean within each layer before scoring candidates. The information actually used is therefore the relative weighting of temporal states within a layer: scaling all weights of a layer merely rescales its objective, whereas a nonconstant temporal profile can change candidate rankings and which states enter the tail term.

We quantified how much these profiles differ across blocks. Each block contributes one mean-normalized temporal profile over the calibration states used for that backbone. DD is the mean pairwise absolute difference between block profiles, averaged over calibration states; "Temporal states" is the number of distinct denoising states; CV is the cross-block coefficient of variation at each state.

ModelBlocksTemporal statesDDCross-block CV range
Wan-1.3B3060.07492.64-19.21%
Wan-14B4060.10274.14-29.90%
MiniMax-H35080.21187.12-27.05%

The profiles are thus block-dependent after normalization, with cross-block variation of up to 19-30% at individual states. We will clarify that Eq. 6 exploits relative temporal sensitivity within a layer, not absolute cross-layer importance.

Q3: Evaluation prompts and seeds.

The evaluation set is the union of the VBench prompt subsets for six dimensions: Imaging Quality (IQ), Aesthetic Quality (AQ), Motion Smoothness (MS), Dynamic Degree (DD), Subject Consistency (SC), and Background Consistency (BC), with duplicates removed, giving 326 unique prompts. We use the expanded prompt descriptions (https://github.com/Vchitect/VBench/blob/master/prompts/augmented_prompts/Wan2.1-T2V-1.3B/all_dimension_aug_wanx_seed42.txt) and keep their dimension assignments. Each prompt is generated once with seed 23001+i23001+i, where ii is the zero-based prompt index; the same prompts and seeds are used for all backbones and methods, and for both the VBench and the paired metrics.

Q4: Differences between Tables 1 and 3/7, and layer coverage.

We apologize for not distinguishing the two settings. The component ablations in Tables 3 and 7 use λ=0.5\lambda=0.5 and ρ=0.25\rho=0.25, whereas the main results use λ=0.75\lambda=0.75 and ρ=0.5\rho=0.5. The PulseQuant entries in Table 1 and Tables 3/7 are therefore different configurations of the same method, which explains the gap (PSNR 15.64 vs. 16.08, IQ 64.96 vs. 65.83).

λ\lambda and ρ\rho are shared across all backbones rather than tuned per model. The shared setting was chosen for its behavior across backbones; it improves some backbones and lowers others. We kept the single shared setting in the main tables instead of reporting the best per-model values. We will state both settings in the table captions and the experimental protocol.

OrbitQuant has no λ\lambda or ρ\rho and quantizes the same modules, so its result is unaffected by this difference: under the protocol of Tables 3/7, it is the Wan-1.3B W4A6 row of Table 1 (IQ 63.78, PSNR 15.30, SSIM 0.521, LPIPS 0.424). PulseQuant is above it on PSNR under both settings (15.64 and 16.08).

"300-layer coverage" is not a separate protocol: it refers to the 300 linear modules in the 30 Transformer blocks of Wan-2.1-T2V-1.3B. Each block contributes ten modules: the query, key, value, and output projections of self-attention and of cross-attention (eight modules), plus the two feed-forward projections. All compared methods, including OrbitQuant, quantize this same set of modules in every table.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

Dear Reviewers,

We sincerely thank all reviewers for your careful evaluations and constructive feedback. Beyond our individual responses, we would like to summarize the key contributions of PulseQuant, the strengths recognized by the reviewers, and the additional experiments conducted during the rebuttal.

[Our Contributions]

  • A diagnostic of local-to-global misalignment. Isolated 4-bit block–timestep interventions show that local reconstruction error is a poor predictor of final latent error, whereas short-horizon propagated error is far more informative.
  • Propagation-guided offline calibration. Propagation-weighted row-radius selection and response-subspace correction with neighboring-code edits, both of which keep the 4-bit weight representation unchanged and add no inference operations.
  • Consistent fidelity gains across five video generators. At W4A4, PulseQuant achieves the best PSNR, SSIM, and LPIPS on all five models, with PSNR gains of 0.16-0.45 dB over OrbitQuant, together with a native packed-NVFP4 deployment path.

[Strengths Recognized by Reviewers]

  • The local-to-global diagnostic is useful and addresses a meaningful limitation of local calibration (KDZz, 2Aq9, bHsV).
  • The method is format-preserving and fully offline, without an additional high-precision branch (KDZz, 2Aq9).
  • The evaluation is broad and up to date, with W4A4 fidelity gains extending across five generators (KDZz, 2Aq9, bHsV).
  • The ablations support the complementary benefits of the two calibration stages (2Aq9, bHsV).
  • The native NVFP4 deployment provides a real deployment path with memory and latency reductions (KDZz, bHsV).

[New Experiments]

We conducted the following additional experiments to address the reviewers' concerns:

  • Radius-search weighting ablation: uniform, timestep-decay, local-error, and the proposed weights, isolating propagation weighting from generic radius search (KDZz).
  • Variance estimates: paired prompt-level bootstrap confidence intervals against OrbitQuant, and calibration stability across calibration prompts and seeds (KDZz).
  • Quality versus deployment numerics: weight- and output-level checks of the NVFP4 export, and end-to-end quality of the native path against Q/DQ simulation (KDZz, 2Aq9).
  • Matched correlation protocol: pooled and within-timestep correlations for all predictors on 256 new interventions, plus an end-to-end comparison of gg-based and EHE_H-based weighting (2Aq9).
  • Cross-block variation of the temporal weight profiles on three backbones (KDZz).
  • Offline calibration cost: runtime, GPU-hours, and peak memory on three models, and scaling with the anchor grid, denoising steps, and horizon HH (KDZz, 2Aq9, bHsV).
  • Correction design space: more protected directions and wider code-index edits (bHsV).
  • Codebook generality: subspace correction under a per-channel uniform INT4 quantizer (bHsV).

We have also clarified the evaluation protocol, hyperparameter settings, and baseline configurations (KDZz). All new results and clarifications will be incorporated into the revised manuscript.

We hope these results and clarifications address the reviewers' concerns. We sincerely appreciate your time and consideration.

Yours sincerely,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.