acceptodds
Under review as a conference paper at ICLR 2027

MatchVR: Bridging the Reference Utilization Gap in Generative Video Restoration

Abstract

Generative video restoration can produce abundant details from severely degraded video, but its strong prior may hallucinate textures inconsistent with the true scene, especially when severe degradation removes reliable high-frequency evidence. Reference-guided restoration reduces this by anchoring reconstruction to real textures from high-quality reference frames, yet reference availability does not imply reference utilization: simply appending reference frames to a video diffusion transformer leaves much of the reference unused, even when reliable correspondences still exist. To address this gap, we present MatchVR, a one-step reference-guided generative video restoration framework that converts reference content into frame-specific matched guidance. Specifically, we introduce a degradation-robust feature matching strategy to estimate dense target-reference correspondences and their confidence. Additionally, we propose a spatiotemporal reference alignment module to warp and relocate the matched content to each target frame. We further introduce aligned-reference distillation to transfer full-reference teacher behavior. Experiments show that MatchVR consistently improves reference utilization and restoration quality across varying temporal distances, with particularly strong gains in fidelity.

Primary area:
generative models
Keywords:
Video Restoration Reference Diffusion model

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_1NtJOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes MatchVR, which reframes reference-guided generative video restoration as a spatiotemporal reference alignment problem. The method first establishes reliable correspondences between low-quality target frames and high-quality reference frames through degradation-robust feature matching. It then resamples the matched reference content into each target frame’s layout in the latent space, and injects it as per-frame local guidance into a DiT backbone via temporal relocation and attention masking. In addition, it distills a full-reference teacher into a single-reference student to improve reference utilization under the single-reference setting.

Critical strengths
  1. The motivation is clear, and the problem is well defined. The authors reveal a “reference utilization gap” through diagnostic experiments: the single-reference curve degrades as the target frame moves away from the reference frame, whereas the idealized full-reference setting remains high and flat. The target-to-reference attention also peaks near the reference frame. This analysis is intuitive and convincing.
  2. The method design is reasonable, and the logical pipeline is complete. Degradation-robust matching, selective spatial alignment, temporal relocation, attention masking, and reference distillation work together to transform a single reference frame into per-frame guidance that is spatially aligned and temporally local.
  3. The efficiency overhead is small. Table 4 shows that adding the aligned reference reduces FPS from 2.77 to 2.70 and increases peak GPU memory from 19.9 GB to 20.1 GB, indicating that the locally visible aligned tokens do not introduce substantial inference overhead.
Critical weaknesses
  1. The diagnosis of the reference utilization gap is not rigorous enough. Figure 1 uses a low-motion, occlusion-free SPMCS subset, but temporal distance also reduces true overlap due to motion, occlusion, and new content. Without reporting matching coverage, occlusion/new-content ratios, and their relation to PSNR, it remains unclear how much of the performance drop comes from underutilization rather than the reference no longer covering the target.
  2. The method relies heavily on high-quality references. Table 3 and Figure 5 show that output quality is strongly correlated with reference quality. The paper does not sufficiently examine the robustness of the method when the reference quality is limited or when the target regions do not appear in the reference.
  3. The paper does not demonstrate the accuracy of LQ matching using DINOv3 features. It would be helpful to compare and visualize LQ–Ref matching against GT–Ref matching to show the accuracy of the current matching stage.
Questions for authors
  1. Can you report matching coverage, the proportion of valid aligned tokens, and occlusion/new-content ratios at different temporal distances, and relate them to PSNR gains? When the aligned reference is empty, does MatchVR degenerate into blind restoration?
  2. What is the essential difference between MatchVR and existing reference-based VSR methods? RefVSR and RefVSR++ already perform reference alignment and propagation. Is MatchVR’s degradation-robust matching plus latent warping merely transferring these mechanisms to a diffusion/DiT backbone? Can you provide a flow-warp or deformable-attention baseline under the same backbone?
5 comments
AuthorsOpenReview ↗

Responses to Reviewer 1NtJ(Part I)

To reviewer 1NtJ: We sincerely thank the reviewer for the careful reading and constructive comments. Responses to specific comments are given below, which are also added to the revised paper.

[W1, Q1]Matching Reliability Analysis

We add the matching statistics at different temporal distances on the low-motion SPMCS subset (GT target frames, η=0.70):

Frame offset relative to the reference frame15102030
Matching coverage0.9960.9910.96420.86920.8138
Valid aligned token ratio0.9960.9910.96420.86920.8138
Matching confidence0.87280.8660.85360.82820.8121
PSNR25.9425.6524.9624.3823.82

The valid aligned token ratio is identical to the coverage under this setting, so the two rows coincide. Even at the largest temporal distance, 81.4% of the target tokens still receive valid matches from the reference under the specified confidence threshold, so the observed quality degradation cannot be attributed to "the reference no longer covering the target". Moreover, the decay rates of the two quantities are clearly mismatched: from frame 1 to frame 30 the coverage decreases by only 0.18, whereas PSNR drops by 2.12 dB, so most of the quality loss cannot be explained by the shrinking matchable area. The statistics and discussion have been added to Section 3 (Motivation) of the revised manuscript (Table 1).

We also add an unrelated-reference experiment on SPMCS:We feed the GT first frame of a different video as the reference. When the reference frame is completely unrelated, the valid matching coverage drops to 6.7 % and the restoration quality smoothly falls back to the level of its own blind restoration, confirming that erroneous reference information is effectively rejected. The results have been added to the Unrelated References paragraph of Section 5.5 "Reference Quality and Robustness Analysis" (Table 7 and Fig. 6(a)).

Reference providedValid matched token ratioPSNR ↑SSIM ↑LPIPS ↓
Correct GT first frame≈92 %25.510.79190.1053
Completely unrelated frame6.70%23.270.61470.4366
No reference (blind restoration)–23.230.60760.4697
AuthorsOpenReview ↗

Responses to Reviewer 1NtJ(Part II)

[W2]Reference-quality Study

We add a 7-level reference-degradation study on SPMCS: across all levels, the PSNR(23.05–25.92 dB) is always higher than SeedVR2-3B (22.19 dB), and using the first frame restored by SeedVR2-3B itself as the reference still gives 22.40 dB. The discussion has been expanded in the Influence of Reference Quality paragraph of Appendix D (Table 15 and Table 16).

Reference sourceReference vs. GT first frame (PSNR / SSIM)Output PSNR ↑Output SSIM ↑Output LPIPS ↓
L0 clean- / -25.510.79190.1053
L1 mild32.31 / 0.927725.920.78230.1406
L2 light29.10 / 0.863625.830.77060.2136
L3 moderate24.89 / 0.673424.230.66350.3641
L4 heavy22.21 / 0.482423.130.57530.4564
L5 severe20.27 / 0.386723.070.58580.4364
L6 extreme18.93 / 0.355223.050.59360.4203
First frame restored by SeedVR221.77 / 0.604222.40.61170.3007
SeedVR2-3B no reference (Table 2)–22.190.63670.3194

We also add a partial-overlap evaluation on 10 high-motion clips from REDS, with the GT first frame as the reference:

MethodReferencePSNR ↑SSIM ↑LPIPS ↓
MatchVRGT first frame23.690.65270.2675
MatchVRNone (blind restoration)22.920.63070.4273
SeedVR2-3BNone23.040.64910.2764
FlashVSR-FullNone21.430.56530.273

The matching coverage decreases monotonically with temporal distance (96.6% at frame 0 down to 19.1% at frame 90), yet under strong motion and partial overlap the reference still brings consistent gains over blind restoration, and remains better than SeedVR2-3B and FlashVSR-Full. The experiment has been added to the Large Motion and Partial Overlap paragraph of Section 5.5 "Reference Quality and Robustness Analysis" in the revised manuscript (Table 8 and Fig. 6(b)).

AuthorsOpenReview ↗

Responses to Reviewer 1NtJ(Part III)

[W3]LQ–GT Matching Robustness

We add Fig. 7 in the updated manuscript to visualize the matching results using GT and degraded target frames. We further quantify matching robustness under degradation using the protocol in Appendix B, where coordinate agreement measures the fraction of matches within one feature cell of the ground-truth correspondences:

DatasetTarget frame typeMatching coverageValid token ratioCoordinate agreement rateconfidence
SPMCSGT (ideal upper bound)0.92040.920410.8428
SPMCSMild0.92510.92510.85390.8442
SPMCSModerate0.90840.90840.76860.836
SPMCSSevere0.81160.81160.66780.8084
YouHQ40GT (ideal upper bound)0.90750.907510.8391
YouHQ40Mild0.90830.90830.84370.8398
YouHQ40Moderate0.89750.89750.72720.8356
YouHQ40Severe0.84680.84680.61540.8173

Under degradation the matching still maintains very high spatial coverage (81.2–92.5 %), and even under severe degradation the coordinate agreement rate remains above 60 %. A lower agreement rate does not necessarily mean wrong matches: in videos the same region often has multiple plausible matching candidates (especially background regions), and the final restored video benefits from the matching results, as corroborated by the full-reference results in Table 2 and Table 6. The comparison has been added to Appendix B.4 "LQ-Ref vs. GT-Ref Matching" of the revised manuscript (Table 11 and Fig. 7).

[Q2]Optical Flow Comparison

This question helps clarify the positioning of our method. RefVSR and RefVSR++ align and propagate reference features within regression-based video super-resolution frameworks. In contrast, MatchVR differs in that: (i) our matcher is explicitly trained to be robust to severe degradations, with confidence-based filtering of unreliable matches; (ii) our alignment aims to improve reference utilization by spatially aligning reference latents, relocating them to each target time step, and restricting their visibility to the corresponding frame; (iii) restoration is performed by a one-step generative DiT. Our key distinction lies in making reliable reference information more accessible to the generative model.

We further add a comparison between optical flow and our matcher under the same evaluation protocol and train a RAFT-based restoration variant by replacing our matcher with RAFT:

DatasetMethodMildModerateSeverePSNRSSIMLPIPS
SPMCSMatchVR0.90230.89290.870225.510.79190.1053
SPMCSRAFT0.97250.74290.63924.540.76940.1127
YouHQ40MatchVR0.90640.89840.882524.940.74580.1360
YouHQ40RAFT0.97560.89910.787724.300.72780.14

Our matcher drops much less in accuracy than optical flow as the degradation becomes more severe, which is what enables the rejection behavior above. Moreover, MatchVR outperforms the RAFT-based variant on all three restoration metrics. The comparison has been added to the Matching Accuracy paragraph of Appendix B.3 (Table 12), and the complete accuracy–coverage results under multiple thresholds are given in Table 9 and Table 10 of Appendix B.

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

The paper studies reference-guided generative video restoration and identifies a "reference utilization gap": when a high-quality reference frame is simply concatenated with target tokens in a video DiT, restoration quality degrades as the target moves away from the reference in time, even when matchable content remains. To address this, MatchVR (built on Wan2.2-TI2V-5B with LoRA, one-step) (i) trains a degradation-robust DINOv3-based matcher to estimate dense LQ-to-reference correspondences and confidence, (ii) warps confidently matched reference latents into each target frame's layout and relocates them to the target's temporal position, visible only to that frame ("aligned reference"), and (iii) trains the model in two phases: a reference-guided baseline, followed by distillation from a frozen full-reference teacher into a single-reference student that uses the aligned reference.

Experiments on SPMCS, YouHQ40, UDM10, VFHQ, and VideoLQ show that, with the ground-truth first frame as reference, MatchVR achieves the best PSNR/SSIM/LPIPS and IDS. Practical variants using ODTSR/PiSA-SR-restored first frames are also reported. Ablations cover the aligned reference (Table 2, Fig. 4), reference quality (Table 3), matcher robustness (Appendix B), and spatial vs. temporal alignment (Appendix C).

Critical strengths
  1. Clear problem diagnosis. The analysis in Sec. 3 / Fig. 1, which separates the "availability ceiling" from the "utilization gap" using per-frame PSNR and target-to-reference attention, is a useful and intuitive framing of why concatenation-based reference guidance under-uses the reference. This motivation is the main reason I find the direction worthwhile.

  2. Consistent gains over the reference-guided baseline under the same reference. Given the same reference source (GT, ODTSR, or PiSA-SR), MatchVR generally outperforms SparkVSR on full-reference metrics, with negligible inference overhead (Table 4). The matcher evaluation in Appendix B, which reports both accuracy and coverage across degradation levels, is also careful and convincing.

Critical weaknesses
  1. The main fidelity claims rely on the GT-reference setting; the advantage over strong blind models in the practical setting is unclear. The headline results (best PSNR/SSIM/LPIPS/IDS) use the ground-truth first frame, which comes from the target video itself and is not available in real applications. In the more realistic setting with an ODTSR/PiSA-SR-restored reference (Table 1), MatchVR does not seem to consistently outperform the blind SeedVR2-3B on fidelity:

    • SSIM is lower on all four datasets (e.g., SPMCS 0.5746/0.5670 vs. 0.6367; UDM10 0.7481/0.7475 vs. 0.7818).
    • PSNR is lower on SPMCS and UDM10, and LPIPS is worse on UDM10 and VFHQ.
    • IDS on VFHQ is lower (0.8338/0.8247 vs. 0.8732), so the identity-preservation claim appears to hold only with GT references.

    The gains in this setting are mainly on no-reference metrics, whereas Appendix D argues that no-reference metrics do not reflect fidelity. Table 3 also suggests that the restored references themselves have lower SSIM than SeedVR2-3B's video output on both SPMCS and YouHQ40 (e.g., 0.5618 vs. 0.6367 on SPMCS), although this compares a single frame with a video average. Interestingly, MatchVR's PSNR advantage over SeedVR2-3B seems to appear only when the reference PSNR itself exceeds SeedVR2-3B's (e.g., YouHQ40 with ODTSR: 23.06 dB reference, 22.96 vs. 22.07 dB output), which suggests that the benefit depends strongly on whether the reference is better than what a blind model already produces. I may be missing something, but the paper does not seem to discuss when a realistically obtainable reference is expected to outperform a strong blind model, which I believe is central to the practical value of the approach.

  2. The ablations may not isolate the contribution of the aligned reference and the distillation. Appendix C states that the LoRA checkpoint is kept fixed across variants, and the "w/o aligned ref" numbers in Table 2 are identical to those in Table 7. If I understand correctly, the "w/o aligned ref" variant is therefore the Phase-2 model (trained with aligned tokens) with these tokens removed at inference. A performance drop in this case may largely reflect a train–test input mismatch rather than the benefit of the design; it is notable that this variant (24.71 dB on SPMCS) is even below SparkVSR with the same GT reference (25.27 dB). The same concern applies to the spatial-only and temporal-only variants in Table 7. In addition, Phase 2 changes both the input (aligned reference) and the objective (distillation instead of a GT loss), and no variant separates the two.

Questions for authors

Clarifications

  1. Phase-2 teacher. My understanding is that the teacher is the frozen Phase-1 model, which receives the LQ video plus the corresponding GT frame as reference for every frame, and the student is supervised only by the teacher's output (Eq. 10), with no GT loss. Is this correct? If so, could you explain why distilling from the teacher is preferable to supervising the student directly with the GT (or combining both)? Since the full-reference teacher itself is not close to GT (Fig. 1a), matching it may cap the student's performance.

  2. Teacher input format. How are the per-frame GT references arranged for the teacher (token layout, positional encoding, attention mask)? Was the Phase-1 model trained with multiple references, or only a single one? If only one, the teacher may operate on an input distribution it was not trained on.

  3. Ablation setup. Is the "w/o aligned ref" variant in Table 2 the same Phase-2 checkpoint with aligned tokens removed at inference (as Appendix C suggests), or a separately trained model? Could you also report the Phase-1 baseline under the single-reference setting?

  4. Fig. 4 attention. Does "attention to reference tokens" in Fig. 4 (left) include the aligned-reference tokens? If so, an increase seems expected once these tokens are added; clarification would help interpret this evidence.

Suggestions (would strengthen the paper)

  1. Independently trained variants under a matched training budget, e.g.: Phase 1 + continued training without aligned reference; Phase 2 with aligned reference but GT loss instead of distillation; and separately trained spatial-only / temporal-only models.

  2. A discussion or experiment on when a practical reference helps over a strong blind model, for example: (a) sweeping reference quality (GT degraded to different levels) to locate the crossover point with SeedVR2-3B; (b) using SeedVR2-3B's own restored first frame as the reference; or (c) a setting with a genuinely external high-quality reference (e.g., multi-camera data).

7 comments
AuthorsOpenReview ↗

Responses to Reviewer XEhD (Part I)

To reviewer XEhD: We sincerely thank the reviewer for the careful reading and constructive comments. Responses to specific comments are given below, which are also added to the revised paper.

[W1, Suggestion 6] Reference-quality Study

We add a systematic reference-quality study with two settings, with results reported below:

1) 7 progressive degradation levels are applied to the GT first frame on SPMCS

2) the first frame restored by SeedVR2-3B itself is additionally used as the reference:

Reference sourceReference vs. GT first frame (PSNR / SSIM)Output PSNR ↑Output SSIM ↑Output LPIPS ↓
L0 clean- / -25.510.79190.1053
L1 mild32.31 / 0.927725.920.78230.1406
L2 light29.10 / 0.863625.830.77060.2136
L3 moderate24.89 / 0.673424.230.66350.3641
L4 heavy22.21 / 0.482423.130.57530.4564
L5 severe20.27 / 0.386723.070.58580.4364
L6 extreme18.93 / 0.355223.050.59360.4203
First frame restored by SeedVR221.77 / 0.604222.40.61170.3007
SeedVR2-3B no reference (Table 2)–22.190.63670.3194

All 8 reference settings outperform SeedVR2-3B in PSNR (+0.21 to +3.73 dB), including the extreme level whose reference is only 18.93 dB; in terms of SSIM the crossover with SeedVR2-3B occurs between L3 and L4, and for LPIPS between L2 and L3. Using the output of SeedVR2-3B itself as the reference, our method still outperforms SeedVR2-3B, showing that it can further extract information from the output of a no-reference model. These results have been added to the Influence of Reference Quality paragraph of Appendix D (Table 15 and Table 16).

3) Real-world high quality reference

We further validate MatchVR on a typical reference-based video restoration scenario, real-world Livephoto videos, in which a genuinely external high-quality reference is available. Livephoto uses separate image and video pipelines: the still image remains sharp and clean, while the video undergoes stronger compression, resulting in softer and noisier content. Without ground truth, we report no-reference metrics:

MethodReferenceARNIQA ↑ILNIQE ↓DOVER ↑
SeedVR2-3B–0.605327.13650.65
FlashVSR-Full–0.61225.7110.68
MatchVRhigh-quality image0.620924.69950.65

MatchVR achieves the best ARNIQA and ILNIQE, consistent with the trade-off analyzed at the end of Appendix D: MatchVR follows the reference evidence rather than the preference of a no-reference predictor**.** Figure 8 shows four restored frames of two clips together with their references, which inherit the reference texture and color while staying temporally coherent. These results are added to Appendix E (Table 18 and Figure 8).

AuthorsOpenReview ↗

Responses to Reviewer XEhD (Part II)

[Q1] GT Loss Ablation

As the reviewer correctly understands, the teacher is the frozen Phase-1 model and the student is supervised only by the teacher's output (Eq. 10) without any GT loss. We add a controlled variant in which the Phase-2 objective is replaced by GT loss while all other conditions are kept exactly the same; The reported metrics are PSNR/SSIM/LPIPS.

Phase-2 objectiveSPMCSYouHQ40
Distillation25.51 / 0.7919 / 0.105324.94 / 0.7458 / 0.1360
GT loss26.01 / 0.7795 / 0.113925.07 / 0.7401 / 0.1403

This controlled comparison has been added to the Distillation Objective paragraph of Section 5.4 and Table 4. The teacher's outputs are jointly guided by high-quality references and the pretrained generative prior, potentially providing supervision that better aligns with the model's reconstruction characteristics. Consistent with this interpretation, GT supervision achieves slightly higher PSNR, while distillation yields better SSIM and LPIPS on both datasets, suggesting that distillation may be more beneficial for preserving structural and perceptual characteristics. We therefore favor distillation for its better balance between reconstruction fidelity and perceptual quality.

[Q2] Teacher Settings

The teacher uses the same model as Phase 1; For a 53-frame clip, for example, it receives 14 GT frames, i.e., each LQ latent is paired with a clean GT frame, and the temporal positional encodings of the reference frames are aligned with those of the corresponding LQ latent. All reference frames are globally visible to all target frames through self-attention. Regarding the training distribution, Phase 1 uses a single reference in 70% of the samples and 2–3 references in 30% of the samples, and the temporal positions of the reference frames are random, so the teacher is capable of using multiple references rather than being restricted to a fully aligned single-reference distribution.

[Q3] Ablation Setup Clarification

The "without aligned reference" variant is not a Phase-2 model with the aligned tokens removed at inference time, but the Phase-1 model itself: it is trained separately, supervised on GT with ℓ1+LPIPS, and contains no aligned reference injection at all, i.e. the Phase-1 baseline under the single-reference setting.

[Q4] Fig. 4 Explanation

Fig. 4 (left) sums the attention over all tokens on the reference side, which includes both the original reference frame latents and all aligned reference latents; This statistical definition is intentional. Since the aligned tokens are derived from the same reference frame via matching, the rise of the curve indicates that the content of this reference is genuinely accessed, rather than that an additional source of information has been introduced.

[W2, Suggestion 5] Spatiotemporal Alignment Ablation

We train three ablation variants—without aligned references, with spatial alignment only, and with temporal relocation only—all initialized from the same Phase-1 model and further trained using the same distillation objective and matched the training budget as the full model. The GT-loss variant and the independently trained model without aligned references are further discussed in our responses to Question 1 and Question 3, respectively. The corresponding ablation results are reported below in PSNR/SSIM/LPIPS:

Datasetw/o aligned reference(Phase1)w/o aligned reference(Phase1+distillation)Spatial alignment onlyTemporal alignment onlyFull aligned reference (MatchVR)
SPMCS24.71 / 0.7739 / 0.115923.41 / 0.7417 / 0.128624.98 / 0.7783 / 0.113424.35 / 0.7400 / 0.130325.51 / 0.7919 / 0.1053
YouHQ4024.36 / 0.7352 / 0.142623.37 / 0.7067 / 0.151624.46 / 0.7329 / 0.149624.09 / 0.7046 / 0.154324.94 / 0.7458 / 0.1360

The full aligned reference still achieves the best restoration quality. Spatial alignment alone increases the amount of usable reference but does not address the attenuation of reference attention caused by temporal distance, so its gain over the variant without aligned reference is limited; temporal alignment alone increases temporal accessibility but cannot guarantee content correspondence. Without aligned references, the student cannot fully benefit from distillation, while the teacher's outputs differ from GT, making distillation less effective than direct GT supervision. Both alignment components are necessary for the final gain.

I thank the authors for the detailed response and the additional experiments. The clarifications on the Phase-2 teacher, the "w/o aligned ref" variant, and Fig. 4 answer my questions. However, after reading the new results, my two main concerns are not yet fully resolved, and I would like to ask the authors to explain the following points.

1. Practical reference setting

  • Where practical references fall. In terms of reference PSNR/SSIM with respect to the GT first frame, the restored references used in practice (ODTSR 21.48 dB / 0.5618, PiSA-SR 21.52 / 0.5464, SeedVR2-3B 21.77 / 0.6042; Tables 6 and 16) all appear to be below L3 (24.89 / 0.6734). At L3, the output LPIPS (0.3641) is already worse than SeedVR2-3B's (0.3194), and from L4 onward the output SSIM is also lower. Moreover, the output PSNR at L4–L6 (23.05–23.13 dB) is close to MatchVR's own no-reference result (23.23 dB, Table 7). My reading is therefore that the PSNR advantage over SeedVR2-3B at these levels comes from MatchVR's blind behavior rather than from the reference. Is this reading correct? If so, the new study seems to indicate that practically obtainable restored references fall near or below the crossover. In which practical situations do the authors expect a restored reference to outperform a strong blind model?

  • Restored references vs. MatchVR's own blind mode. On SPMCS, adding an ODTSR / PiSA-SR reference lowers PSNR and SSIM compared with MatchVR without a reference (21.74 / 0.5746 and 21.93 / 0.5670 vs. 23.23 / 0.6076; Table 2 vs. Table 7), while LPIPS improves substantially (0.2933 / 0.2985 vs. 0.4697). This seems to suggest that a restored reference mainly transfers sharper details that are not necessarily faithful to the GT. That would run counter to the paper's motivation of anchoring restoration to real textures to reduce hallucination. How do the authors explain this result? Is there evidence that, in this setting, the details MatchVR transfers from the reference are faithful rather than hallucinated content inherited from the reference?

2. Isolating the effect of the aligned reference

  • The new distillation control. It falls below Phase 1 (23.41 vs. 24.71 dB on SPMCS). The improvement of the full model over this control may therefore partly reflect that distillation without aligned reference is harmful, rather than that the aligned reference is beneficial.
  • The remaining comparisons. Each of them changes more than one factor:
    • Phase 1 (24.71 dB) vs. the full model: the full model differs in the aligned reference, the training objective, and the additional Phase-2 training.
    • Phase 1 vs. the GT-loss variant with aligned reference (26.01 dB, Table 4): this still differs in both the aligned reference and the additional training.
  • Question. How do the authors separate the contribution of the aligned reference from that of the additional training? Is there evidence that continued GT-loss training of Phase 1 without the aligned reference, under the same budget, would not close this gap?

I look forward to the authors' response.

Reviewer_7dVmOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents MatchVR, a reference-guided generative video restoration framework that uses a single high-quality reference frame to guide the restoration of an entire video. The method consists of three main components: a degradation-robust feature matcher, selective spatial alignment with temporal relocation, and a two-stage reference distillation strategy. The matcher identifies reliable correspondences between the reference and target frames, while the alignment mechanism makes the matched reference features more accessible to the restoration model. The distillation strategy further enables the single-reference student to learn from a full-reference teacher. Experiments on multiple video restoration benchmarks show improvements in reconstruction fidelity

Critical strengths

1.The method addresses several related challenges in reference-guided video restoration, including correspondence estimation, spatial alignment, temporal accessibility, and reference utilization. 2. Experiments on multiple benchmarks show consistent improvements in PSNR, SSIM, and LPIPS when aligned references are introduced, supporting the effectiveness of the proposed design.

Critical weaknesses

1.The matching evaluation mainly uses image pairs with known geometric transformations and synthetic degradations. It does not systematically examine unrelated references, misleading high-similarity matches, or substantial scene changes. Consequently, it remains unclear how reliably the method rejects incorrect reference information in challenging real-world scenarios. 2.Temporal relocation improves the accessibility of reference features but does not explicitly enforce consistent correspondences across adjacent frames. The experiments mainly report reconstruction and perceptual quality metrics, without directly evaluating temporal flickering or cross-frame correspondence stability. Therefore, the method's effectiveness in maintaining temporal consistency is not fully established. 3.The results show a noticeable performance gap between ground-truth references and references restored by existing methods. While this limitation is acknowledged by the authors, the experiments do not fully characterize the method's robustness to severely degraded, partially incorrect, or poorly matched references. This limits the conclusions that can be drawn about its practical performance when high-quality references are unavailable.

Questions for authors

1.How does MatchVR behave when the reference frame is unrelated to the target video or contains visually similar but incorrect content? Can the confidence threshold reliably reject such matches, and what happens to restoration quality when matching coverage becomes low? 2.Does temporal relocation improve temporal consistency in addition to increasing reference attention? Could the authors provide temporal consistency metrics or cross-frame matching visualizations to demonstrate that the method does not introduce flickering or unstable textures? 3.How does performance change when the reference contains only partial overlap with the target video, or when substantial new content appears in later frames? Can the method identify when reference guidance is unreliable and fall back to restoration without reference guidance? 4.How does MatchVR's end-to-end inference speed compare with competing methods under the same hardware and resolution settings? How does its performance scale to longer videos and more substantial motion or scene changes? 5.Could the authors provide additional comparisons under identical reference conditions to better isolate the benefits of the proposed matching and alignment mechanisms from those of simply having access to an additional high-quality reference?

5 comments
AuthorsOpenReview ↗

Responses to Reviewer 7dVm (Part I)

To reviewer 7dVm: We sincerely thank the reviewer for the careful reading and constructive comments. Responses to specific comments are given below, which are also added to the revised paper.

[W1, Q1] Unrelated Reference

We add an unrelated-reference experiment on SPMCS: We feed the GT first frame of a different video as the reference:

Reference providedValid matched token ratioPSNR ↑SSIM ↑LPIPS ↓
Correct GT first frame≈92 %25.510.79190.1053
Completely unrelated frame6.70%23.270.61470.4366
No reference (blind restoration)–23.230.60760.4697

When the reference frame is completely unrelated, the valid matching coverage drops to 6.7 % and the restoration quality returns to nearly the same level as its own no-reference restoration baseline, confirming that erroneous reference information is effectively rejected instead of corrupting the output. MatchVR itself also has blind restoration capability: its PSNR is slightly better than the compared blind restoration method but not on perceptual metrics such as LPIPS, consistent with its design goal of exploiting high-quality references rather than optimizing blind restoration. The results have been added to the Unrelated References paragraph of Section 5.5 "Reference Quality and Robustness Analysis" (Table 7 and Fig. 6(a)).

2.[W2, Q2] Temporal Consistency Metric

We add a temporal consistency evaluation using tLPIPS over adjacent frame pairs to the full evaluation set. The detailed computation of tLPIPS is provided in the Metrics paragraph of Section 5.2 of the revised manuscript.

MethodReferenceSPMCS ↓YouHQ40 ↓
MatchVR (Spatial + Temporal)GT first frame0.053450.07209
MatchVR (Spatial only)GT first frame0.053810.0725
MatchVR(blind)-0.061980.08781
SparkVSRGT first frame0.056720.10122
SeedVR2-3B–0.065180.07965
Vivid-VR–0.067880.09659
FlashVSR-Full–0.082080.10739
STAR–0.082430.09932

MatchVR is the temporally most stable among all compared methods on both datasets. Compared with its blind-restoration variant, MatchVR achieves lower tLPIPS, indicating that reference guidance further enhances temporal consistency. We note that the direct effect of Temporal Relocation on tLPIPS is small; its main contribution is making the aligned reference genuinely accessible to the frame it belongs to, while temporal stability benefits from both the one-step latent modeling of the backbone and reference guidance. The results have been added to the tLPIPS rows of Table 2 (all four datasets), and the conclusions of Section 5.3 have been updated accordingly.

AuthorsOpenReview ↗

Responses to Reviewer 7dVm (Part II)

[W3, Q3] Different Reference Robustness

1) Partial-overlap Evaluation

We add a large-motion / partial-overlap evaluation on 10 high-motion clips from REDS, with the GT first frame as the reference.

MethodReferencePSNR ↑SSIM ↑LPIPS ↓
MatchVRGT first frame23.690.65270.2675
MatchVRNone (blind restoration)22.920.63070.4273
SeedVR2-3BNone23.040.64910.2764
FlashVSR-FullNone21.430.56530.273

We add matching coverage (reference - each video frame) curve in updated Fig.6. The matching coverage decreases monotonically with temporal distance (96.6% at frame 0 down to 19.1% at frame 90), yet under strong motion and partial overlap the reference still brings consistent gains over blind restoration, and remains better than SeedVR2-3B and FlashVSR-Full.

2)Reference-quality Study

We add a systematic reference-quality study with two settings, 7 progressive degradation levels are applied to the GT first frame on SPMCS and the first frame restored by SeedVR2-3B itself is additionally used as the reference:

Reference sourceReference vs. GT first frame (PSNR / SSIM)Output PSNR ↑Output SSIM ↑Output LPIPS ↓
L0 clean- / -25.510.79190.1053
L1 mild32.31 / 0.927725.920.78230.1406
L2 light29.10 / 0.863625.830.77060.2136
L3 moderate24.89 / 0.673424.230.66350.3641
L4 heavy22.21 / 0.482423.130.57530.4564
L5 severe20.27 / 0.386723.070.58580.4364
L6 extreme18.93 / 0.355223.050.59360.4203
First frame restored by SeedVR221.77 / 0.604222.40.61170.3007
SeedVR2-3B no reference (Table 2)–22.190.63670.3194

In addition, across all 7 reference-degradation levels on SPMCS, the output PSNR stays within 23.05–25.92 dB and is always higher than SeedVR2-3B, indicating a wide tolerance to reference quality.

The two detailed analyses have been added to the Large Motion and Partial Overlap paragraph of Section 5.5 "Reference Quality and Robustness Analysis" (Table 8 and Fig. 6(b)) and to the Influence of Reference Quality paragraph of Appendix D (Table 15 and Table 16).

AuthorsOpenReview ↗

Responses to Reviewer 7dVm (Part III)

[Q4] Efficiency and Long Video

We add an efficiency comparison under identical hardware and resolution settings (720p, a single A100), and a long-video evaluation on 10 clips of 200–600 frames using a single reference frame. MatchVR runs at 2.70 FPS with aligned references, while SeedVR2-3B measures 1.06 FPS. Most long clips are captured with stable cameras without drastic motion, and the average PSNR/SSIM/LPIPS is 27.84/0.8318/0.1954. MatchVR is therefore clearly faster than SeedVR2-3B and extends directly to long videos with a single reference. The comparison has been added to the Efficiency Analysis paragraph of Section 5.4 and Table 5, and the long-video results to the Long Videos paragraph of Appendix E and Table 17.

[Q5]Additional Comparisons

We add a controlled comparison that fixes the reference condition (the same GT first frame) and changes only the alignment mechanism:

With the reference condition fixed, MatchVR improves all metrics on both datasets, so the gain comes from the matching and alignment mechanism. The main ablation is presented in the Aligned Reference paragraph of Section 5.4 and Table 3, and the complete decomposition including the spatial-only and temporal-only rows in Appendix C and Table 13.The reported metrics are PSNR/SSIM/LPIPS

Datasetw/o aligned referenceSpatial alignment onlyTemporal alignment onlyFull aligned reference (MatchVR)
SPMCS24.71 / 0.7739 / 0.115924.98 / 0.7783 / 0.113424.35 / 0.7400 / 0.130325.51 / 0.7919 / 0.1053
YouHQ4024.36 / 0.7352 / 0.142624.46 / 0.7329 / 0.149624.09 / 0.7046 / 0.154324.94 / 0.7458 / 0.1360

Comments on the paper

1 comment
Area_Chair_aUh8OpenReview ↗

Reviewer-author discussion period is ending. Please feedback.

Dear Reviewers,

Thanks for your time and effort in reviewing ICLR2027 submissions.

As the reviewer-author discussion period will end very soon, please take time to read the author’s rebuttal and give your feedback. Your continuous engagement is highly appreciated.

Best regards,

Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.