acceptodds
Under review as a conference paper at ICLR 2027

Temporal Re-Distillation: A Post-Distillation Stabilizer for One-Step Video Super-Resolution

14 views

Abstract

One-step video super-resolution (VSR) models distilled via Distribution Matching Distillation (DMD) can preserve the frame-wise perceptual quality of multi-step diffusion teachers, yet often exhibit pronounced temporal artifacts, including flickering and texture drift. To understand this discrepancy, we trace temporal refinement along the teacher's denoising trajectory and observe that fine-grained cross-frame consistency is refined predominantly at low noise levels, a regime largely bypassed by one-step distillation. Motivated by this observation, we propose TRD-VSR, a temporal refinement framework that re-perturbs student predictions to low-noise states and distills the teacher's local denoising corrections in this regime. We further introduce Relational Structure Alignment (RSA) to explicitly transfer cross-frame relational structure, complementing the local teacher corrections with an explicit constraint on temporal correspondence. TRD-VSR is used only during training and therefore preserves single-step inference with no additional test-time cost. Experiments on three benchmarks show that TRD-VSR improves both temporal consistency and perceptual quality, achieving the best tLPIPS across all datasets, the best tOF and tLP on two datasets, and the best CLIPIQA and DOVER across all datasets, yielding a favorable perception–consistency trade-off over existing one-step generative VSR methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
video restore, video super-resolution, diffusion model, one-step distillation, generative model, 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_ye1GOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies temporal artifacts in a one-step DMD video super-resolution student. It analyzes how a 50-step teacher's attention becomes more localized along motion correspondences at low noise, then proposes TRD-VSR: low-noise correction distillation plus relational structure alignment to HR-video feature relations. The extra teacher and HR supervision are training-only, preserving one-step inference. Component ablations on UDM10, SPMCS, and YouHQ40 show improved tOF and Ewarp over the starting student, and the method obtains the best tLPIPS and DOVER among listed methods on all three benchmarks.

Critical strengths
  • The method keeps one denoiser evaluation at inference and uses a matched one-step starting student for LNCD and RSA ablations. On UDM10, their combination reduces tOF from 1.384 to 0.870 and Ewarp from 0.877 to 0.821, with the same trends appearing on the other two main benchmarks.
  • The authors look beyond a single temporal score: tOF, tLPIPS, tLP and Ewarp are considered, and the low-noise attention intervention gives some diagnostic support. The additional MVSR4x benchmark and measured inference time/memory are useful checks.
Critical weaknesses
  • The attention-based explanation remains suggestive rather than decisive. Figure 2 shows an association between denoising step and attention locality, Appendix B.2 tests only 10 UDM10 clips and blocks neighborhoods across all 30 teacher layers. Blocking within-frame neighborhoods also harms temporal metrics. A matched intervention at high noise or a less disruptive control is needed to isolate late cross-frame refinement as the cause.
  • The perceptual/temporal gains come with notable reference-fidelity tradeoffs: UDM10 PSNR is 22.91 versus DOVE's 25.36 and YouHQ40 PSNR is 21.24 versus DOVE's 23.21. The claim of generally better temporal consistency needs this mixed evidence stated more prominently.
Questions for authors
  1. Can the authors compare low-noise correction to an equal-compute high-noise correction and to continued training with the same extra 5K updates, to isolate the benefit of the proposed mechanism?
  2. What explains the much worse MVSR4x Ewarp despite improved tLPIPS, and do human observers prefer the resulting motion consistency under real degradations?
5 comments
AuthorsOpenReview ↗

Response to Reviewer ye1G [Critical Weaknesses 1]

Critical Weaknesses 1: Attention analysis and matched interventions

Reviewer concern: A matched intervention at high noise or a less disruptive control is needed to isolate late cross-frame refinement as the cause.

Matched attention-masking experiments across noise levels.

To address the reviewer's concern about the lack of matched high-noise interventions, we conducted additional experiments across three noise regimes (Table R1), using 30 clips: all 10 UDM10 clips and 10 clips randomly selected with a fixed seed from each of SPMCS and YouHQ40.

For each clip, we first generated an unblocked reference using the teacher's full 50-step denoising trajectory. We then repeated the trajectory from the same initial noise, applying the same cross-frame attention mask for exactly four consecutive steps at three stages: low noise (steps 46–49, σ≤0.30\sigma \leq 0.30), intermediate noise (steps 42–45, 0.32≤σ≤0.470.32 \leq \sigma \leq 0.47), and high noise (steps 0–3, σ≈1\sigma \approx 1). The mask prevents each token from attending to nearby spatial positions in neighboring frames. Thus, the masking operation and intervention duration remain fixed, while only the denoising stage varies. EwarpE_{\mathrm{warp}} and tLPIPS are normalized per clip to the corresponding unblocked trajectory and reported as median [IQR]; Δ\DeltaPSNR and Δ\DeltaLPIPS are mean per-clip differences.

Table R1. Blocking four steps of the teacher trajectory across three datasets (n=30n=30). Superscripts denote paired Wilcoxon signed-rank tests against the unblocked trajectory: ∗∗p<0.01^{**}p<0.01 for degradation and †p=0.001^{\dagger}p=0.001 for improvement.

InterventionEwarp↓E_{\mathrm{warp}}\downarrowtLPIPS↓ΔPSNR↑ΔLPIPS↓
Unblocked trajectory1.001.000.000.000
Cross-frame (σ ≤ 0.30)2.00 [1.60, 3.35]**1.40 [1.30, 1.99]**−0.550.021
Cross-frame (0.32 ≤ σ ≤ 0.47)1.24 [1.09, 1.94]**1.20 [1.12, 1.66]**−0.170.017
Cross-frame (σ ≈ 1)0.90 [0.78, 1.00]†1.00 [0.99, 1.02]+0.15−0.001
Within-frame (σ ≤ 0.30)1.41 [1.25, 1.74]**1.60 [1.31, 1.72]**−0.300.045

Results.

Blocking cross-frame attention causes the largest temporal degradation at low noise (normalized EwarpE_{\mathrm{warp}}: 2.00; tLPIPS: 1.40), smaller degradation at intermediate noise (1.24; 1.20), and no corresponding degradation at high noise (0.90; 1.00). The ordering in EwarpE_{\mathrm{warp}} is consistent across all three datasets. These matched interventions provide stronger evidence that cross-frame attention plays a particularly important role in temporal refinement during late denoising.


Why does masking within-frame neighbors also reduce temporal stability?

At low noise, masking within-frame neighbors worsened both temporal metrics and caused a larger increase in single-frame LPIPS than masking cross-frame neighbors (+0.045 vs. +0.021, Table R1). These results suggest distinct but complementary roles for the two interactions, consistent with the increasingly localized spatiotemporal attention in Fig. 2. Within-frame neighbors provide spatial context for resolving fine textures, while cross-frame neighbors help maintain texture alignment across motion, reducing frame-to-frame shifts that appear as flicker. We will clarify this interpretation in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer ye1G [Critical Weaknesses 2]

Critical Weaknesses 2: Reference fidelity and temporal consistency

Reviewer concern: The perceptual/temporal gains come with notable reference-fidelity tradeoffs.

Fidelity and temporal consistency.

Table 2 shows that our temporal gains hold at comparable reference fidelity. FlashVSR [1] is the closest method to ours in PSNR: the two methods differ by at most 0.15 dB on each of the three datasets. Our method achieves higher SSIM on all three and performs better in eight of the nine temporal comparisons. These results show that we can improve temporal consistency while maintaining similar fidelity.

Why our PSNR and SSIM are lower than those of some methods.

The comparison with FlashVSR shows that our temporal gains do not require a substantial loss of fidelity. The lower PSNR and SSIM relative to some other methods are consistent with the established perception–distortion trade-off [2]. VSR is an ill-posed inverse problem with multiple plausible solutions. Minimizing mean squared error to maximize PSNR tends to favor smooth outputs that average over plausible reconstructions. Distortion metrics can therefore penalize visually plausible details that differ from the particular ground-truth reference [3]. Lower PSNR or SSIM thus need not imply lower perceptual quality. Like Upscale-A-Video, STAR, DLoRAL, and FlashVSR [1, 4–6], we aim to recover realistic, detailed videos. Our work further addresses the temporal instability of those generated details. Table 2 illustrates this distinction: DOVE [7], SeedVR2 [8], and PS-SR [9] achieve higher PSNR than our method, while our method scores higher on the no-reference perceptual quality metrics.

Overall results.

In Table 2, our method ranks first in 7 of the 9 temporal comparisons and 8 of the 9 no-reference perceptual quality comparisons, while maintaining reference fidelity comparable to FlashVSR. These results support our central contribution of improving both perceptual quality and temporal consistency at competitive reference fidelity.

References

[1] Zhuang et al. FlashVSR: Towards Real-time Diffusion-Based Streaming Video Super Resolution. CVPR, 2026.

[2] Blau and Michaeli. The Perception-Distortion Tradeoff. CVPR, 2018.

[3] Ledig et al. Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network. CVPR, 2017.

[4] Zhou et al. Upscale-A-Video: Temporal-Consistent Diffusion Model for Real-World Video Super-Resolution. CVPR, 2024.

[5] Xie et al. STAR: Spatial-Temporal Augmentation with Text-to-Video Models for Real-World Video Super-Resolution. ICCV, 2025.

[6] Sun et al. One-Step Diffusion for Detail-Rich and Temporally Consistent Video Super-Resolution. NeurIPS, 2025.

[7] Chen et al. DOVE: Efficient One-Step Diffusion Model for Real-World Video Super-Resolution. NeurIPS, 2025.

[8] Wang et al. SeedVR2: One-Step Video Restoration via Diffusion Adversarial Post-Training. ICLR, 2026.

[9] Wu et al. PS-SR: Pseudo-Single-Step Video Super-Resolution via Speculative Diffusion. CVPR, 2026.

AuthorsOpenReview ↗

Response to Reviewer ye1G [Questions 1]

Q1. Equal-compute teacher correction and continued DMD training

Can the authors compare low-noise correction to an equal-compute high-noise correction and to continued training with the same extra 5K updates, to isolate the benefit of the proposed mechanism?

We conducted the following controlled comparisons to test whether the gains come from low-noise correction, additional training, or standard teacher distillation.

Shared settings.

All post-training variants start from the same one-step DMD VSR student checkpoint and use the same training data, optimizer, learning rate, and 5K additional generator updates. For the comparison across correction noise levels, we also fix the number of teacher evaluations and vary only the starting noise level in Eq. (8). We evaluate all variants on UDM10, SPMCS, and YouHQ40. The baseline is the DMD checkpoint before the additional 5K updates.

ExperimentSettingPurpose
OursApply low-noise correction at σL=0.15\sigma_L=0.15 and train with the objective in Eqs. (8)–(12).Proposed method.
(a) Continued DMD trainingContinue training with the original DMD objective in Eq. (5).Test whether 5K additional updates alone explain the gains.
(b) Higher-noise correctionKeep the same correction procedure and training objective as ours, changing only the starting noise level to σL=0.35\sigma_L=0.35 or 0.50.5.Test whether the gains depend on performing correction at low noise.
(c) Standard teacher distillationStarting from pure Gaussian noise, the teacher generates a target over multiple denoising steps conditioned on the LR input. The student learns this target through endpoint regression in Eq. (9).Test whether standard teacher distillation yields comparable gains.

UDM10 (n = 10)

VarianttOF ↓EwarpE_{\mathrm{warp}} ↓tLPIPS ↓CLIPIQA ↑MUSIQ ↑DOVER ↑
Baseline1.3840.8770.11250.56664.40.591
Ours (σL=0.15\sigma_L=0.15)0.8700.8210.10400.57266.40.621
Higher-noise correction (σL=0.5\sigma_L=0.5)0.9750.9190.11320.58266.50.594
Standard distillation (σ=1\sigma=1)1.0560.5920.14460.50258.00.448
DMD + 5K updates1.2630.9010.11340.54263.90.602
Higher-noise correction (σL=0.35\sigma_L=0.35)0.9350.9510.11280.57666.30.601

SPMCS (n = 30)

VarianttOF ↓EwarpE_{\mathrm{warp}} ↓tLPIPS ↓CLIPIQA ↑MUSIQ ↑DOVER ↑
Baseline0.3720.7910.05730.57471.00.552
Ours (σL=0.15\sigma_L=0.15)0.3330.7850.05760.57470.30.555
Higher-noise correction (σL=0.5\sigma_L=0.5)0.3611.1310.06240.58171.20.520
Standard distillation (σ=1\sigma=1)0.3770.7180.08120.50166.10.371
Higher-noise correction (σL=0.35\sigma_L=0.35)0.3641.0530.06100.58371.40.533
DMD + 5K updates0.3590.8280.05980.56570.60.543

YouHQ40 (n = 40)

VarianttOF ↓EwarpE_{\mathrm{warp}} ↓tLPIPS ↓CLIPIQA ↑MUSIQ ↑DOVER ↑
Baseline1.8101.6020.11570.56065.50.709
Ours (σL=0.15\sigma_L=0.15)1.1491.4660.10760.55866.40.716
Higher-noise correction (σL=0.5\sigma_L=0.5)1.3751.7340.11280.56067.30.719
Standard distillation (σ=1\sigma=1)1.6090.7490.12510.43757.80.635
Higher-noise correction (σL=0.35\sigma_L=0.35)1.1271.6580.11180.56166.90.723
DMD + 5K updates1.7081.5160.11930.54667.20.707

Results.

Continuing DMD training for 5K updates improves tOF over the initial checkpoint on all three datasets, but worsens tLPIPS on all three and yields higher errors than our method on all three temporal metrics. With the number of teacher evaluations held fixed, raising the starting noise level from 0.15 to either 0.35 or 0.5 increases both EwarpE_{\mathrm{warp}} and tLPIPS on every dataset. This supports the benefit of performing teacher correction at low noise. Standard distillation from pure noise achieves the lowest EwarpE_{\mathrm{warp}}, but also the worst tLPIPS and the lowest CLIPIQA, MUSIQ, and DOVER scores on all three datasets, indicating a substantial loss of the perceptual quality achieved by the DMD student.

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

This paper addresses temporal inconsistency in one-step video super-resolution (VSR) models distilled using Distribution Matching Distillation (DMD). The authors observe that multi-step diffusion teachers progressively refine temporal correspondences, particularly at low noise levels, which are not fully preserved in one-step students. To address this, they propose Temporal Re-Distillation (TRD-VSR), combining Low-Noise Correction Distillation (LNCD) and Relational Structure Alignment (RSA) to improve temporal consistency while maintaining single-step inference.

Critical strengths
  1. The paper addresses an important limitation of aggressive diffusion distillation: preserving frame-level perceptual quality does not necessarily guarantee temporal consistency. The analysis of attention patterns along the teacher's denoising trajectory provides an interesting perspective on how temporal information may be progressively refined.
  2. The proposed method improves an existing one-step student through an additional training stage without modifying its inference architecture.
Critical weaknesses
  1. A central claim is that temporal consistency is recovered primarily through progressive attention refinement in the low-noise regime. However, the presented attention analysis mainly establishes a correlation between attention localization and denoising progress. More controlled experiments are needed to isolate the specific contribution of low-noise refinement from the general benefits of additional teacher denoising.
  2. While the ablation variants share the same training budget, the original DMD student is not compared against an equally trained baseline with additional fine-tuning. Consequently, it remains unclear how much improvement originates from the proposed low-noise correction mechanism rather than additional optimization or more conventional distillation objectives.
Questions for authors
  1. Would it be possible to compare teacher corrections performed at different noise levels while controlling for the number of teacher evaluations? This would help determine whether the improvement specifically originates from the low-noise regime or from additional denoising in general.
  2. The ablation results suggest that RSA does not consistently improve all metrics. For example, LNCD alone achieves better DOVER on UDM10 and YouHQ40, and slightly better tLPIPS on YouHQ40, compared with the full method. These results raise questions about the specific benefits and trade-offs introduced by RSA.
  3. In particular, the proposed method achieves strong tLPIPS but substantially worse Ewarp compared with several baselines. What types of temporal artifacts are captured differently by these metrics, and under which conditions does the proposed approach perform less effectively?
4 comments
AuthorsOpenReview ↗

Response to Reviewer XhsE [Critical Weaknesses 1&2, Question 1]

Response to W1, W2, and Q1. Low-noise correction and matched training controls

W1: “More controlled experiments are needed to isolate the specific contribution of low-noise refinement from the general benefits of additional teacher denoising.”

W2: “Consequently, it remains unclear how much improvement originates from the proposed low-noise correction mechanism rather than additional optimization or more conventional distillation objectives.”

Q1: “Would it be possible to compare teacher corrections performed at different noise levels while controlling for the number of teacher evaluations?”

To address W1 and Q1, we compare teacher correction at different noise levels under matched teacher evaluations. For W2, we additionally compare continued DMD training and conventional teacher distillation. All variants start from the same DMD student checkpoint and use identical training data, optimizer, learning rate, and 5K generator updates. The baseline denotes the checkpoint before these additional updates.

ExperimentSetting and purpose
OursLow-noise correction at σL=0.15\sigma_L=0.15 using Eqs. (8)–(12).
(a) Continued DMDTrain with Eq. (5) for 5K more updates to test the effect of additional optimization.
(b) Higher-noise correctionChange only σL\sigma_L in Eq. (8) to 0.350.35 or 0.50.5, keeping teacher evaluations and training objective fixed.
(c) Conventional distillationGenerate teacher targets from pure noise and train the student with endpoint regression (Eq. 9).

UDM10 (n = 10)

VarianttOF ↓EwarpE_{\mathrm{warp}} ↓tLPIPS ↓CLIPIQA ↑MUSIQ ↑DOVER ↑
Baseline1.3840.8770.11250.56664.40.591
Ours (σL=0.15\sigma_L=0.15)0.8700.8210.10400.57266.40.621
(a) DMD + 5K updates1.2630.9010.11340.54263.90.602
(b) σL=0.35\sigma_L=0.350.9350.9510.11280.57666.30.601
(b) σL=0.5\sigma_L=0.50.9750.9190.11320.58266.50.594
(c) Conventional distillation1.0560.5920.14460.50258.00.448

SPMCS (n = 30)

VarianttOF ↓EwarpE_{\mathrm{warp}} ↓tLPIPS ↓CLIPIQA ↑MUSIQ ↑DOVER ↑
Baseline0.3720.7910.05730.57471.00.552
Ours (σL=0.15\sigma_L=0.15)0.3330.7850.05760.57470.30.555
(a) DMD + 5K updates0.3590.8280.05980.56570.60.543
(b) σL=0.35\sigma_L=0.350.3641.0530.06100.58371.40.533
(b) σL=0.5\sigma_L=0.50.3611.1310.06240.58171.20.520
(c) Conventional distillation0.3770.7180.08120.50166.10.371

YouHQ40 (n = 40)

VarianttOF ↓EwarpE_{\mathrm{warp}} ↓tLPIPS ↓CLIPIQA ↑MUSIQ ↑DOVER ↑
Baseline1.8101.6020.11570.56065.50.709
Ours (σL=0.15\sigma_L=0.15)1.1491.4660.10760.55866.40.716
(a) DMD + 5K updates1.7081.5160.11930.54667.20.707
(b) σL=0.35\sigma_L=0.351.1271.6580.11180.56166.90.723
(b) σL=0.5\sigma_L=0.51.3751.7340.11280.56067.30.719
(c) Conventional distillation1.6090.7490.12510.43757.80.635

Effect of correction noise level (W1, Q1). Under matched teacher evaluations, σL=0.15\sigma_L=0.15 consistently achieves lower EwarpE_{\mathrm{warp}} and tLPIPS than σL=0.35\sigma_L=0.35 or 0.50.5 across all three datasets. This confirms that the benefits depend on the correction noise level, rather than merely on additional teacher denoising. The specific role of attention remains an interpretation supported by our attention analysis.

Additional training and conventional distillation (W2). Compared with continued DMD training, our method achieves lower tOF, EwarpE_{\mathrm{warp}}, and tLPIPS on all three datasets under the same 5K-update budget. Conventional distillation yields lower EwarpE_{\mathrm{warp}} but substantially worse tLPIPS and perceptual quality (CLIPIQA, MUSIQ, DOVER), suggesting that its low warping error comes at the expense of perceptual fidelity. These controls support the distinct benefit of low-noise correction beyond additional optimization or conventional distillation.

AuthorsOpenReview ↗

Response to Reviewer XhsE [Question 2]

Response to Q2. What benefits and trade-offs does RSA introduce?

Q2: “The ablation results suggest that RSA does not consistently improve all metrics. For example, LNCD alone achieves better DOVER on UDM10 and YouHQ40, and slightly better tLPIPS on YouHQ40, compared with the full method.”

RSA mainly adds motion-related structural constraints. Adding RSA on top of LNCD lowers EwarpE_{\mathrm{warp}} by a further 8.4%, 7.5%, and 4.3% on UDM10, SPMCS, and YouHQ40, respectively. It also improves tOF on all three datasets, cutting it by 17.0% on YouHQ40.

On the other hand, DOVER evaluates overall video quality, while tLPIPS measures frame-to-frame perceptual feature distance. LNCD aligns the teacher’s generative prior, whereas RSA encourages its cross-frame feature correlations to match those derived from the GT video. The slight misalignment between these two objectives may limit their additive gains, although RSA alone or LNCD alone improves over the baseline. In Eq. (12), we offer λRSA\lambda_{\mathrm{RSA}} to control this trade-off.

AuthorsOpenReview ↗

Response to Reviewer XhsE [Question 3]

Response to Q3. Differences between warping error and perceptual temporal consistency

Q3: “In particular, the proposed method achieves strong tLPIPS but substantially worse Ewarp compared with several baselines. What types of temporal artifacts are captured differently by these metrics, and under which conditions does the proposed approach perform less effectively?”

Different aspects of temporal consistency. EwarpE_{\mathrm{warp}} and tLPIPS measure temporal consistency from different perspectives. EwarpE_{\mathrm{warp}} computes pixel-wise residuals between motion-aligned adjacent frames and is therefore sensitive to local texture variations, color fluctuations, and contrast changes. In contrast, tLPIPS evaluates temporal differences in perceptual feature space and is more tolerant of small pixel-level variations. Consequently, generative VSR methods may achieve strong perceptual temporal consistency while exhibiting larger warping errors, particularly when restoring high-frequency textures that are not perfectly aligned across frames.

Effect of color alignment on MVSR4x. This discrepancy is particularly evident on MVSR4x, where LR and HR videos are captured by different cameras, introducing additional color and contrast differences absent from synthetic benchmarks. Several existing methods, including FlashVSR, SeedVR2, UltraVSR, DLoRAL, MGLD-VSR, and STAR, employ AdaIN or wavelet-based color correction after inference. Our standard pipeline does not include dataset-specific color alignment, and we originally reported results without this post-processing step.

To isolate its impact, we applied frame-wise AdaIN color alignment to the same restored videos. As shown below, EwarpE_{\mathrm{warp}} decreased substantially from 1.508 to 0.933, whereas tLPIPS increased from 0.1049 to 0.1205. This opposite trend demonstrates that the two metrics respond differently to color processing, and that our originally higher warping error on MVSR4x is partly attributable to the absence of color alignment.

MVSR4x results with and without color alignment

MethodtOF ↓EwarpE_{\mathrm{warp}} ↓tLPIPS ↓PSNR ↑SSIM ↑LPIPS ↓CLIPIQA ↑MUSIQ ↑DOVER ↑
SeedVR2-7B1.6970.5190.157122.570.7670.4100.24031.60.181
DOVE2.3540.5500.111921.900.7530.4020.40262.20.515
FlashVSR2.1711.0750.114921.830.7240.3870.43568.00.544
UltraVSR3.3631.6840.161920.860.7160.4310.47266.20.501
PS-SR2.0810.9930.129521.610.7400.4130.45367.50.487
DLoRAL3.3240.9480.183522.100.7340.4320.50661.40.409
Ours, without color alignment1.7491.5080.104919.090.7040.4110.43967.90.564
Ours, with AdaIN color alignment1.8080.9330.120521.570.7310.4010.42767.10.534

Limitations and perceptual quality trade-off. Our method can be less favorable under pixel-sensitive temporal evaluation, particularly when real-world degradations introduce substantial color discrepancies or when generated fine textures exhibit small frame-to-frame variations. Moreover, overly smooth reconstructions may naturally achieve lower warping errors by suppressing high-frequency details, without necessarily providing better perceptual quality.

After color alignment, our method achieves the third-lowest EwarpE_{\mathrm{warp}} on MVSR4x. Only SeedVR2 and DOVE obtain lower warping errors, but both perform worse on all three no-reference perceptual quality metrics (CLIPIQA, MUSIQ, and DOVER). Among the remaining methods, TRD-VSR achieves the lowest warping error while maintaining strong perceptual quality. These results highlight the importance of evaluating temporal consistency jointly with perceptual quality, rather than relying on a single pixel-based metric.

Thank you for the detailed response. The authors have addressed my concerns, and I will maintain my original score.

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

The paper studies temporal artifacts in one-step video super-resolution obtained through distribution matching distillation. It observes increasing motion-aligned attention concentration during the teacher's low-noise denoising stages and proposes a post-distillation procedure, TRD-VSR. Low-Noise Correction Distillation (LNCD) trains the student against teacher-corrected endpoints and intermediate features, while Relational Structure Alignment (RSA) matches cross-frame feature relations to those of paired HR video. All additional branches are training-only, preserving the original one-step inference architecture. Experiments cover three standard benchmarks, an additional real-world benchmark, component ablations, and inference efficiency.

Critical strengths
  1. Useful temporal improvements without additional inference steps. Table 3 reduces tOF from 1.384 to 0.870 on UDM10 and from 1.810 to 1.149 on YouHQ40, while improving warping error on all three benchmarks. The teacher and auxiliary branches are training-only. This combination supports the method's practical value for efficient generative restoration.
  2. Evidence extends beyond a single benchmark or qualitative examples. The study includes component ablations, a real-world benchmark, runtime/memory measurements, and attention-blocking interventions with paired tests (Appendix B.2). These strengthen the case that the observed temporal improvements warrant further investigation.
Critical weaknesses
  1. The proposed explanation of DMD's temporal weakness is not sufficiently established. Eq. (4) defines a score-difference gradient, but Table 1 reports an underdefined scalar “DMD loss”; similar surrogate values do not necessarily imply similar gradients against temporal perturbations. The fake score network is also frozen while evaluating interleaved samples from a changed distribution. This diagnoses particular trained estimators, rather than establishing that DMD generally optimizes primarily frame-wise quality. Moreover, the attention-blocking intervention is conducted only at σ=0.15, so it does not establish the relative importance of late versus early correction. These gaps matter because the diagnosis is a central contribution motivating the method.
  2. The distinctive benefit of low-noise re-distillation is not isolated from other training effects. Table 3 separates LNCD and RSA, but not endpoint supervision from feature alignment, and does not compare matched-budget correction at another noise schedule. The improvements are also metric-dependent: SPMCS tLPIPS changes from 0.0573 to 0.0576, and the full model does not improve every YouHQ40 metric over LNCD alone. Paired uncertainty for the main comparisons and a focused correction control are needed to establish which part of the proposed training procedure is responsible for the benefit. The existing significance tests concern attention blocking, not these method comparisons.
Questions for authors
  1. Clarification: What is the exact scalar quantity reported as DMD loss in Table 1, including its normalization and noise sampling? Does a noise-level or gradient-sensitive analysis support the objective-level claim, or should that claim be limited to the frozen estimators tested?
  2. Targeted validation: Under the existing post-distillation budget, how do endpoint-only correction and a matched alternative noise schedule compare with LNCD? This would directly test whether low-noise correction, rather than additional teacher-guided training, is the distinctive source of the gain.
  3. Related work and contribution clarification: Could you cite and discuss InstaRevive: One-Step Image Enhancement via Dynamic Score Matching (ICLR 2025) and TSD-SR: One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution (CVPR 2025)? InstaRevive dynamically controls the perturbation range to improve score-based distribution matching, while TSD-SR addresses unreliable teacher guidance and uses distribution-aware sampling to improve access to detail-oriented gradients. How does LNCD differ from these approaches in its objective, noise selection, and teacher supervision? In particular, what additional insight or mechanism is needed for temporal consistency beyond image-level restoration? This clarification would help assess the novelty of the low-noise correction component.
5 comments
AuthorsOpenReview ↗

Response to Reviewer rpob[Critical Weaknesses 1 and Question 1]

Response to W1 and Q1: DMD Loss, Score-Difference Direction, and Noise-Level Intervention

W1: "Eq. (4) defines a score-difference gradient, but Table 1 reports an underdefined scalar 'DMD loss' [...] The fake score network is also frozen [...] Moreover, the attention-blocking intervention is conducted only at σ=0.15."

Q1: "What is the exact scalar quantity reported as DMD loss in Table 1, including its normalization and noise sampling? Does a noise-level or gradient-sensitive analysis support the objective-level claim [...]?"

Definition of the DMD loss in Table 1.

We thank the reviewer for pointing out this ambiguity. Eq. (4) specifies an implicit objective's gradient, while Table 1 reports the normalized squared magnitude of the score difference as a diagnostic proxy:

SDMD(z0)=12DEσ,η[∥w(sreal(zσ,σ)−sfake(zσ,σ))∥22]S_{\mathrm{DMD}}(z_0) = \frac{1}{2D} E_{\sigma,\eta} [\|w(s_{\mathrm{real}}(z_\sigma,\sigma)-s_{\mathrm{fake}}(z_\sigma,\sigma))\|_2^2]

where DD is the latent dimensionality, ww is the DMD2 per-sample normalization, and σ\sigma follows the training noise distribution. We evaluate V1V_1 and VflickerV_{\mathrm{flicker}} using matched noise and the frozen real/fake score networks at the end of DMD training. We acknowledge that Table 1 diagnoses these specific estimators rather than establishing a general limitation of the DMD objective, and will qualify the claim in Sec. 3.3 accordingly.

Direction of the score difference.

To examine whether the score difference captures the teacher's temporal correction, we define δ=zrep−zstudent\delta=z_{\mathrm{rep}}-z_{\mathrm{student}} using the teacher-corrected output from Eq. (8) (σL=0.15\sigma_L=0.15, M=1M=1). We evaluate zα=zrep+α(zstudent−zrep)z_\alpha=z_{\mathrm{rep}}+\alpha(z_{\mathrm{student}}-z_{\mathrm{rep}}) at α∈0,0.5,1\alpha\in\\{0,0.5,1\\}, from the teacher-corrected output to the student output, and measure

C(α,σ)=cos⁡(sreal(zσ,σ)−sfake(zσ,σ),δ),C(\alpha,\sigma)=\cos\left(s_{\mathrm{real}}(z_\sigma,\sigma)-s_{\mathrm{fake}}(z_\sigma,\sigma),\delta\right),

where zσz_\sigma is the noised latent derived from zαz_\alpha.

Table R1. Score-difference alignment with the teacher correction, averaged over 10 UDM10 videos.

α\alphaσ=0.12\sigma=0.12σ=0.20\sigma=0.20σ=0.50\sigma=0.50
0 (Teacher)0.2930.2600.160
0.50.2470.2130.144
1 (Student)0.1760.1570.127

At the student output (α=1\alpha=1), the alignment remains limited (C=0.127C=0.127–0.1760.176). This suggests that the tested frozen score estimators provide only weak directional alignment with the teacher's low-noise correction, motivating LNCD as complementary supervision.

Matched attention masking across noise levels.

We further extend the original σ=0.15\sigma=0.15 intervention to three stages of the teacher's 50-step trajectory, using 30 clips (10 each from UDM10, SPMCS, and YouHQ40). With identical initial noise and attention masks, we block four consecutive steps at low (46–49), intermediate (42–45), or high (0–3) noise, varying only the intervention stage.

Table R2. Four-step attention masking (n=30n=30). EwarpE_{\mathrm{warp}} and tLPIPS are per-clip normalized ratios (median [IQR]); Δ\DeltaPSNR and Δ\DeltaLPIPS are mean differences. Paired Wilcoxon test vs. unblocked: ∗∗p<0.01^{**}p<0.01 (degradation), †p=0.001^{\dagger}p=0.001 (improvement).

InterventionEwarp↓E_{\mathrm{warp}}\downarrowtLPIPS↓ΔPSNR↑ΔLPIPS↓
Unblocked1.001.000.000.000
Cross-frame (σ≤0.30\sigma \le 0.30)2.00 [1.60, 3.35]**1.40 [1.30, 1.99]**−0.550.021
Cross-frame (0.32≤σ≤0.470.32 \le \sigma \le 0.47)1.24 [1.09, 1.94]**1.20 [1.12, 1.66]**−0.170.017
Cross-frame (σ≈1\sigma \approx 1)0.90 [0.78, 1.00]†1.00 [0.99, 1.02]+0.15−0.001
Within-frame (σ≤0.30\sigma \le 0.30)1.41 [1.25, 1.74]**1.60 [1.31, 1.72]**−0.300.045

Cross-frame masking causes the greatest temporal degradation at low noise, a smaller effect at intermediate noise, and no degradation at high noise. The EwarpE_{\mathrm{warp}} ordering holds across all three datasets. Within-frame masking also degrades temporal metrics but causes a larger single-frame quality loss (Δ\DeltaLPIPS: 0.045 vs. 0.021), suggesting distinct contributions from spatial and cross-frame attention. These results support the importance of late-stage cross-frame refinement, and we will incorporate them into the revision.

AuthorsOpenReview ↗

Response to Reviewer rpob[Critical Weaknesses 2 and Question 2] (Part 1)

1. Endpoint supervision versus feature alignment (C2, Q2)

Reviewer concern: "Table 3 separates LNCD and RSA, but not endpoint supervision from feature alignment."

Response. We added an endpoint-only ablation on UDM10. The LNCD row uses the same results as Table 3 in the original paper.

ConfigurationtOF ↓Ewarp ↓tLPIPS ↓CLIPIQA ↑MUSIQ ↑DOVER ↑
DMD2 Baseline1.3840.8770.11250.56664.40.591
Endpoint only (LendL_{\mathrm{end}})0.8910.8090.10630.54865.00.615
Endpoint + feature (LNCD)0.8710.8960.10740.57966.00.625

Endpoint supervision improves all three temporal metrics over the DMD2 baseline. Adding feature alignment further reduces tOF (0.891 → 0.871) and improves CLIPIQA, MUSIQ, and DOVER, although Ewarp and tLPIPS increase relative to endpoint-only supervision. Thus, the endpoint loss provides a substantial part of the temporal benefit, while feature alignment contributes additional tOF and perceptual-quality gains rather than uniformly improving all metrics.

2. Matched noise schedules and additional-training controls (C2, Q2)

Reviewer concern: "The distinctive benefit of low-noise re-distillation is not isolated from other training effects." A matched-budget correction at another noise schedule and controls for additional optimization are needed.

Response. We compare correction at different noise levels with matched teacher evaluations, along with continued DMD training and conventional teacher distillation. All variants start from the same DMD student checkpoint and use the same training data, optimizer, learning rate, and 5K generator updates. “Baseline” denotes the checkpoint before these extra updates.

VariantPurpose / setting
OursLow-noise correction at σL=0.15\sigma_L=0.15 using Eqs. (8)–(12)
(a) DMD + 5KContinue DMD optimization using Eq. (5), isolating extra training updates
(b) Higher-noise correctionSet σL=0.35\sigma_L=0.35 or 0.50.5 in Eq. (8), with matched teacher evaluations and the same training objective
(c) Conventional distillationGenerate teacher targets from pure noise and use endpoint regression (Eq. 9)

UDM10 (n = 10)

VarianttOF ↓Ewarp ↓tLPIPS ↓CLIPIQA ↑MUSIQ ↑DOVER ↑
Baseline1.3840.8770.11250.56664.40.591
Ours (σL=0.15\sigma_L=0.15)0.8700.8210.10400.57266.40.621
(a) DMD + 5K1.2630.9010.11340.54263.90.602
(b) σL=0.35\sigma_L=0.350.9350.9510.11280.57666.30.601
(b) σL=0.5\sigma_L=0.50.9750.9190.11320.58266.50.594
(c) Conventional distillation1.0560.5920.14460.50258.00.448

SPMCS (n = 30)

VarianttOF ↓Ewarp ↓tLPIPS ↓CLIPIQA ↑MUSIQ ↑DOVER ↑
Baseline0.3720.7910.05730.57471.00.552
Ours (σL=0.15\sigma_L=0.15)0.3330.7850.05760.57470.30.555
(a) DMD + 5K0.3590.8280.05980.56570.60.543
(b) σL=0.35\sigma_L=0.350.3641.0530.06100.58371.40.533
(b) σL=0.5\sigma_L=0.50.3611.1310.06240.58171.20.520
(c) Conventional distillation0.3770.7180.08120.50166.10.371

YouHQ40 (n = 40)

VarianttOF ↓Ewarp ↓tLPIPS ↓CLIPIQA ↑MUSIQ ↑DOVER ↑
Baseline1.8101.6020.11570.56065.50.709
Ours (σL=0.15\sigma_L=0.15)1.1491.4660.10760.55866.40.716
(a) DMD + 5K1.7081.5160.11930.54667.20.707
(b) σL=0.35\sigma_L=0.351.1271.6580.11180.56166.90.723
(b) σL=0.5\sigma_L=0.51.3751.7340.11280.56067.30.719
(c) Conventional distillation1.6090.7490.12510.43757.80.635

Interpretation. Continued DMD training yields smaller tOF gains than our method on all three datasets despite the same extra 5K updates. Under matched teacher evaluations, σL=0.15\sigma_L=0.15 achieves lower Ewarp and tLPIPS than either higher-noise correction on all three datasets. Its tOF is also better on UDM10 and SPMCS, whereas σL=0.35\sigma_L=0.35 is slightly better on YouHQ40. Conventional distillation produces lower Ewarp but markedly worse tLPIPS and perceptual quality. Together, these controls support the importance of the low-noise correction regime for a favorable temporal–perceptual balance.

AuthorsOpenReview ↗

Response to Reviewer rpob[Critical Weaknesses 2 and Question 2] (Part 2)

1. Metric-dependent trade-offs (C2)

Reviewer concern (Critical Weakness 2): "The improvements are also metric-dependent: SPMCS tLPIPS changes from 0.0573 to 0.0576, and the full model does not improve every YouHQ40 metric over LNCD alone."

Response. We agree that the improvements are metric-dependent. As shown in Table 3, adding RSA to LNCD further reduces Ewarp by 8.4%, 7.5%, and 4.3% on UDM10, SPMCS, and YouHQ40, respectively, and also improves tOF across all three datasets. However, LNCD alone achieves slightly better DOVER on UDM10 and YouHQ40, and tLPIPS on YouHQ40. On SPMCS, the full model slightly increases tLPIPS from 0.0573 to 0.0576 compared with the baseline.

LNCD aims to align the student's outputs with the teacher's generative prior, whereas RSA encourages the generated video's cross-frame feature correlations to match those of the GT video. The slight misalignment between these objectives may limit their additive gains across metrics, although both LNCD and RSA individually provide improvements over the baseline on several metrics. We introduce the weighting coefficient λRSA\lambda_{\mathrm{RSA}} in Eq. (12) to balance these objectives.

2. Paired uncertainty of the main method comparison (C2)

Reviewer concern (Critical Weakness 2): "Paired uncertainty for the main comparisons and a focused correction control are needed to establish which part of the proposed training procedure is responsible for the benefit. The existing significance tests concern attention blocking, not these method comparisons."

Response. We conducted paired video-level analysis comparing the DMD baseline with the full model on all three datasets. tOF improves on 9/10, 25/30, and 30/40 videos on UDM10, SPMCS, and YouHQ40, respectively. The paired 95% confidence intervals for tOF exclude zero on UDM10 ([−53.5%, −17.7%]) and YouHQ40 ([−58.5%, −5.8%]), with p < 0.01 for both comparisons. The paired 95% confidence intervals for tLPIPS likewise exclude zero on UDM10 ([−12.1%, −2.5%]) and YouHQ40 ([−9.3%, −4.9%]).

These results provide statistical support for the baseline-to-full-model temporal improvements on those datasets, while we do not claim statistical significance for unreported comparisons or metrics.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.