Response to Reviewer svnJ (denoted as R4)

Q4-1 Despite the simplicity and intuitiveness, the originality is limited. The DOVE is simply adapted from commonly used perceptual VSR/ISR losses. It barely uses the abundant diffusion priors from CogVideoX except for initialization.

A4-1 Thank you for the comment. We clarify our novelty and prior utilization below.

(1) Novelty Clarification.

1. Common Loss. The commonly used loss functions, such as MSE, DISTS, and Frame losses, are not our novelty. We adopt these commonly used losses to ensure general applicability.

2. (Novelty) Training Strategy. Our main contribution lies in the proposed latent-pixel training strategy, which consists of two stages. Stage-1 learns the one-step mapping in the latent space. Stage-2 enhances reconstruction in pixel-space. Additionally, we introduce image-video mixed training and a frame-by-frame VAE encoder-decoder to better adapt to VSR training.

3. Experiment Results. Table 1a in the main paper demonstrates the effectiveness of our training strategy. We provide the results below:

   Training Stage	Stage-1	Stage-1+Stage-2(Image)	Stage-1+Stage-2(Image/Video)
   PSNR↑	27.20	26.39	26.48
   LPIPS↓	0.3037	0.2784	0.2696
   CLIP-IQA↑	0.3236	0.5085	0.5107
   DOVER↑	0.6154	0.7694	0.7809

4. (Another Novelty) Data Pipeline. We also design a video processing pipeline to construct the HQ-VSR dataset. This dataset also contributes to the performance.

(2) Diffusion Prior.

1. Prior Utilization. We apply pretrained diffusion priors for initialization to enhance the restoration capability of DOVE.
2. Task Gap. Since the original model (i.e., CogVideoX) is designed for generation rather than restoration, we propose the latent-pixel training strategy to better adapt it to the VSR task.
3. Effective Training Design. Our design ensures architectural simplicity, training efficiency, and impressive performance.

Q4-2 The visual results tend oversharp and smooth.

A4-2 Thank you for the comment. In some challenging cases, the visual results of our DOVE may appear oversharpened and smooth. However, this typically occurs in highly degraded scenarios where existing methods struggle to recover the right structure.

We provide several examples in the supplementary material to illustrate this point.

• Case 1 (Fig. 5, first sample, UDM10: 000): The LR input is heavily blurred, and the wall textures are barely visible. Our DOVE successfully reconstructs the correct structural lines, although slightly oversharpened. In comparison, other methods fail to recover the lines and even introduce artifacts.
• Case 2 (Fig. 7, fourth sample, VideoLQ: 016): While DOVE produces a slightly oversharpened result, it restores the text correctly and without artifacts. In contrast, methods like STAR misrecognize the word "CRIPPLE" as "CHIPPLE", while other methods produce artifacts.

Overall, DOVE achieves superior restoration quality compared to existing methods. Of course, we acknowledge the reviewer's observation that some challenging cases may exhibit oversharpening or smoothing. This is an important direction for future improvement.

Q4-3 My main question is about the utilization of diffusion priors. Since the DOVE is simply adapted from commonly used perceptual VSR/ISR losses and does not use CogVideoX to calibrate the output, I have doubts about identifying the DOVE as a "diffusion" model. For instance, replacing the CogVedeo with a GAN, like GigaGAN to VideoGigaGAN, the pipeline still works. It seems like the improvement is brought by the scaling laws of changing the base model from weak GigaGAN to powerful CogVideoX.

A4-3 Thank you for the question. Our approach is a diffusion model. We explain it below.

(1) Following the Diffusion Paradigm.

1. Diffusion Formulation. DOVE follows the core principle of diffusion models: generating results through denoising. Specifically, our model adopts the v-prediction scheme, where a Transformer predicts the velocity (a variant of noise). The restored latent ${z}_{sr}$ is obtained using:

   $z_{sr} = \sqrt{ᾱ_t} z_{lr} - \sqrt{1-ᾱ_t} v_θ(z_{lr}, t)$.

   This formulation is consistent with the definition of the diffusion model.

2. Consistency with Existing Methods. Our method aligns with the framework of other one-step diffusion models (e.g., those discussed by Reviewer-WLko in Q3-1: UltraVSR [1], DLoRAL [2], and OSEDiff [3]). From the framework definition perspective, the main difference between these diffusion methods and our DOVE is the loss function, but this does not affect the overall definition.

(2) Pipeline Transferability.

1. The pipeline is transferable. One strength of our method is its generalizability. The proposed training strategy can be adapted to other pretrained T2V backbones, not just CogVideoX.

2. Although pipeline transfer is possible, diffusion priors are central to our approach. Compared to GANs, diffusion models offer more stable training and stronger generative priors. These properties contribute significantly to the impressive performance of DOVE.

3. Motivated by the reviewer's comment, we extend DOVE to another pretrained diffusion T2V model, WAN2.1-T2V-1.3B [ref1]. We refer to the resulting model as DOVE (WAN2.1). The evaluation results on UDM10 are as follows:

   Method	PSNR↑	LPIPS↓	CLIP-IQA↑	DOVER↑
   Upscale-A-Video	21.72	0.4116	0.4697	0.7291
   MGLD-VSR	24.23	0.3272	0.4557	0.7264
   STAR	23.47	0.4242	0.2417	0.4830
   DOVE (WAN2.1)	25.28	0.2555	0.5133	0.7563

   The WAN-based model shows strong VSR performance, validating the generality and effectiveness of our method.

[ref1] Wan: Open and Advanced Large-Scale Video Generative Models, 2025.

Q4-4 In Table 1 (b), the DOVE (80%) leads by DOVE (100%) in a rather small margin. However, Table 2 compares only ResShift, which is a rather weak baseline even for image SR. Comparison with more recent diffusion-based SR and GAN-based VSR will raise my confidence in DOVE.

A4-4 Thank you for the feedback. We provide detailed clarifications below.

(1) Improvement Clarification. Compared with DOVE (100%), DOVE (80%) has a quantitative improvement of 0.0115 in the DOVER metric, which is not very small. Meanwhile, image-video mixed training leads to more stable and consistent results.

(2) More Comparison Methods. We include comparisons with more recent methods:

• SinSR [ref2]: Image SR, diffusion-based, CVPR 2024
• OSEDiff [ref3]: Image SR, diffusion-based, NeurIPS 2024
• RealViformer [ref4]: Video SR, GAN-based, ECCV 2024

(Note: VideoGigaGAN is not publicly released, so we cannot include it in the evaluation.)

The comparison results are as follows:

The results show that DOVE outperforms recent diffusion- and GAN-based methods in video quality and temporal consistency.

[ref2] SinSR: Diffusion-Based Image Super-Resolution in a Single Step, CVPR 2024.

[ref3] One-Step Effective Diffusion Network for Real-World Image Super-Resolution, NeurIPS 2024.

[ref4] RealViformer: Investigating Attention for Real-World Video Super-Resolution, ECCV 2024.

Q4-5 The visual results in sup tend to be extremely sharp and lose details, making it look unnatural. Moreover, the high sharpness and smoothness will induce higher subjective scores.

A4-5 Thank you for the comment.

In challenging cases with severe degradation, DOVE may produce results that appear oversharpened or overly smooth. However, it still outperforms existing methods regarding fidelity and visual quality.

Additionally, we agree with the reviewer that better balancing sharpness and detail preservation is a valuable direction for future work.

For more details, please see A4-2.

Response to Reviewer WLko (denoted as R3)

Q3-1 Insufficient novelty. One-step diffusion models were proposed in many existing methods, i.e., [1]-[4]. The difference between the proposed DOVE and existing methods should be clarified and discussed. Additionally, in video processing pipeline, the metadata/scene/quality filtering were proposed by existing methods, i.e., Stable video diffusion, Openvid-1m, Open-sora. The DOVE only added more metrics, which is insufficient novelty.

A3-1 Thank you for the comment. We clarify the novelty of our one-step model and data pipeline below.

(1) One-Step Novelty.

General Difference.

1. OSEDiff [3] and SinSR [4] are designed for image SR and are not tailored to the video task. Meanwhile, UltraVSR [1] and DLoRAL [2] are one-step VSR methods released on arXiv after the NeurIPS submission deadline.
2. Existing methods [1]–[4] rely on pretrained image models and incorporate specifically designed modules to enhance performance.
3. In contrast, our method is built upon a pretrained video model. We propose a well-designed training strategy that activates and enhances the VSR ability of the pretrained model.

Detailed Difference.

1. UltraVSR [1] enhances temporal consistency by inserting temporal layers into a pretrained image model (i.e., stable diffusion) and uses a degradation-aware module to improve performance.
2. DLoRAL [2] also relies on the image model (stable diffusion) uses a dual-stage process to train consistency-LoRA and detail-LoRA modules. However, it only supports two-frame inference.
3. OSEDiff [3] is an image SR method using LoRA and variational score distillation, leveraging pretrained image priors.
4. SinSR [4] distills a multi-step image SR model (i.e., ResShift) into a one-step student within the image domain.
5. DOVE (ours) utilizes a pretrained video model and introduces the latent-pixel training strategy designed explicitly for VSR, achieving high-quality one-step restoration.

Comparison Results.

We compare DOVE against existing one-step SR methods: OSEDiff, SinSR, and DLoRAL (UltraVSR code is not publicly available). The comparison results are as follows:

We observe that image SR methods (OSEDiff and SinSR) perform poorly on video-specific metrics, e.g., DOVER or $E^*_{warp}$, confirming their unsuitability for VSR tasks. Compared to DLoRAL, our DOVE achieves superior performance, validating the effectiveness of our method.

(2) Pipeline Novelty.

1. Metadata Filtering Difference. Unlike previous pipelines, our process emphasizes high-resolution video, which is essential for VSR and not explicitly focused on in other pipelines.
2. Quality Filtering Difference. As the eviewer (WLko) noted, we incorporate more diverse quality metrics. This is consistent with the goal of VSR, which is to recover fine-grained details.
3. Motion Filtering Difference. This step is unique to our pipeline. This fits the characteristics of VSR clipping training and is not included in previous pipelines.
4. Ablation Results. Table 1d in the main paper demonstrates the effectiveness of the filtering and motion process. Further ablation for each filter component is provided in A3-2.
5. Comparison with Openvid-1M. As shown in Tab. 1c (main paper), models trained on HQ-VSR outperform those trained on the existing dataset, Openvid-1M. This demonstrates the value of our video processing pipeline.

Q3-2 Experimental results. In Table 1-3 provide the corresponding results that proves the effectiveness of DOVE model. Some detailed results should be provided. For example, The image ratio interval in Table 1 (b) is too large, and setting a smaller interval is beneficial for obtaining the optimal ratio. The experimental results of each filter operation in video processing pipeline should be provided in Table 1(d) for verifying the effectiveness of each filter. Besides, in Table 1 (d), the "+filter" (steps 2-4) only brings a PSNR gain of 0.05dB. How to verify the effectiveness of the proposed filters?

A3-2 Thank you for the suggestions. We conduct additional ablations for the image ratio and each filtering step. We also clarify the filter gains.

(1) Ablation: Fine-Grained Image Ratio.

We conduct more fine-grained experiments on the image ratio. Based on the preliminary results in Tab. 1b, we compare the cases of 70%, 75%, 80%, 85%, and 90%, and the results are as follows:

We find that 80% and 85% yield a good balance between frame fidelity and perceptual quality. This is close to the settings (80%) in our paper. For consistency, we still adopt the image ratio of 80%.

(2) Ablation: Data Filtering Steps.

We perform a breakdown ablation of the three filtering components in our data pipeline: Metadata, Scene, and Quality filtering. The results are as follows:

Different filters have different functions:

• Metadata filtering removes low-resolution videos that are unsuitable for VSR. This improves fidelity (e.g., PSNR) and slightly enhances perceptual quality.
• Scene filtering has a limited effect since OpenVid-1M already includes scene cuts. However, it is essential to extend the pipeline to other unprocessed data.
• Quality filtering significantly enhances perceptual metrics, particularly DOVE. The slight drop in PSNR may result from imperfect alignment between restored details and the ground truth. But it still performs better than the unfiltered dataset (i.e., OpenVid-1M).

(3) Clarification on Filter Effectiveness.

1. Caption Correction: The label, "+Filter (steps 2–4)" in Tab. 1 (main paper) is a typo. Actually, it is "+Filter (steps 1-3)". We will revise it.
2. Filter-Specific Effects: As shown above, each filter contributes differently to fidelity and perceptual performance.
3. Perceptual Improvement: The primary purpose of filtering is to improve perceptual quality. While the overall PSNR gain is 0.05 dB, perceptual improvements are significant. For example, DOVER improves by 0.0994.

Q3-3 Comparison of Fairness: DOVE model proposed the video processing pipeline to fine-tuning the VSR models on high-quality datasets. In state-of-the-art VSR comparison in Table 2, the DOVE model using the proposed video processing pipeline achieves the best performance in most scenarios, while other methods without used. Can other existing methods also use this pipeline to improve performance? The more results and discussion should be provided.

A3-3 Thank you for the comment.

1. Our training strategy is one of the innovations in our method. Therefore, we directly compare DOVE with existing approaches that do not use our training strategy.

2. The training strategy of existing diffusion VSR methods is tightly bound to model design. Thus, directly applying our training strategy to those models is not easy and requires extensive adaptation.

3. But our method is well-suited for pretrained text-to-video (T2V) models. To reveal the generalizability of our method, we conduct experiments using another pretrained T2V model, WAN2.1-T2V-1.3B [ref1], with the same training settings as DOVE. We refer to the resulting model as DOVE (WAN2.1).

   The evaluation results on UDM10 are listed below:

   Method	PSNR↑	LPIPS↓	CLIP-IQA↑	DOVER↑
   Upscale-A-Video	21.72	0.4116	0.4697	0.7291
   MGLD-VSR	24.23	0.3272	0.4557	0.7264
   STAR	23.47	0.4242	0.2417	0.4830
   DOVE (WAN2.1)	25.28	0.2555	0.5133	0.7563

   The WAN-based model also achieves strong VSR performance, further confirming the generalizability of our proposed method.

[ref1] Wan: Open and Advanced Large-Scale Video Generative Models, 2025.

Dear Review WLko,

We sincerely appreciate the time and effort you devoted to reviewing our work, as well as your thoughtful and insightful comments. We have carefully considered your feedback and provided detailed responses to your concerns:

1. We clarify the novelty of our training strategy and dataset pipeline.
2. We conduct more fine-grained ablations on the image ratio and data filtering components.
3. We extend our method to other backbones to demonstrate its generalizability.

As the author–reviewer discussion phase is nearing its conclusion, we would like to confirm whether our responses have addressed your concerns. If you have any questions or suggestions, please do not hesitate to let us know.

Thank you again for your thoughtful feedback and kind support.

Best regards,

Authors

Response to Reviewer Xx99 (denoted as R2)

Q2-1 The only weakness is related to novelty. L_S1 and L_S2 have elements that have already been applied. In the case of L_S1, comparing the latent representations between the prediction and the ground truth is not totally new (i.e., StableSR or DiffBIR). For the case of L_S2, this stage incorporates established components like pixel-space MSE loss and perceptual loss. [1] uses the differences between frames.

A2-1 Thank you for the comment. We clarify our novelty as follows:

1. Loss Function. The specific loss functions, such as MSE, DISTS, and Frame losses, are not claimed to be innovative. We adopt these standard losses to ensure general applicability.

2. (Novelty) Training Strategy. Our core contribution is the proposed latent-pixel training strategy. Unlike common latent-space losses (e.g., StableSR, DiffBIR), our approach enhances restoration in the pixel space. Additionally, we introduce image-video mixed training and the frame-by-frame VAE encoder-decoder to better adapt to VSR training.

3. Experiment Results. Table 1a in the main paper demonstrates the effectiveness of our training strategy. We further show the results here.

   Training Stage	Stage-1	Stage-1+Stage-2(Image)	Stage-1+Stage-2(Image/Video)
   PSNR↑	27.20	26.39	26.48
   LPIPS ↓	0.3037	0.2784	0.2696
   CLIP-IQA↑	0.3236	0.5085	0.5107
   DOVER↑	0.6154	0.7694	0.7809

4. (Another Novelty) Data Pipeline. Another contribution is our video processing pipeline, through which we construct the HQ-VSR dataset. The dataset also contributes to the performance.

In general, while individual loss functions are standard, the overall training strategy, which is adapted for VSR, is novel.

Q2-2 Table 1c provides an initial indication of the significance of the proposed HQ-VSR dataset. My question focuses on two key aspects: how much the performance of the proposed model degrades without training or adaptation using HQ-VSR, and to what extent related work or baseline methods might improve if they were evaluated using the same dataset.

A2-2 Thank you for the insightful question. We conduct two new experiments: (1) train our DOVE on the REDS dataset; (2) retrain the recent method (MGLD-VSR) with our proposed HQ-VSR dataset.

(1) DOVE Trained on REDS.

We train DOVE using the REDS dataset with the same training settings described in our paper. The stage-2 image dataset is still DIV2K. We refer to the resulting model as DOVE (REDS). The results on UDM10 are shown below:

These results show that training on HQ-VSR boosts both fidelity and perceptual quality.

Meanwhile, thanks to the proposed training strategy and pretrained model prior, DOVE (REDS) still outperforms existing methods when trained on REDS.

(2) MGLD-VSR Trained on HQ-VSR.

We retrain MGLD-VSR using its official codebase, replacing the original training dataset (REDS) with HQ-VSR.

We use 4 A6000 GPUs with a batch size of 16. Other training settings are consistent with the official configuration.

The resulting model is denoted as MGLD-VSR (HQ-VSR). The performance on UDM10 is:

Training MGLD-VSR on HQ-VSR also results in improvements across all metrics, demonstrating the general utility of our HQ-VSR.

Q2-3 L_S2 has three components (MSE, Dists, and frame). How does the model perform if one or two elements are removed?

A2-3 Thank you for the suggestion. We conduct an ablation study on the loss functions used in stage-2. The results are as follows:

We observe that MSE alone has poor perceptual performance, although PSNR is high. This may be because the model tends to output smoother results. The DISTS loss significantly improves perceptual quality, as reflected in LPIPS, CLIP-IQA, and DOVER scores. Meanwhile, the Frame loss also brings certain performance improvements by enhancing frame consistency.

Response to Reviewer 1Gww (denoted as R1)

Q1-1 Evaluation bias: Key no-reference metrics (CLIP-IQA, FasterVQA) are also used in dataset filtering, raising concerns about selection bias.

A1-1 Thank you for the comment.

The metrics in our dataset pipeline are general and diverse, covering both frame-level quality (CLIP-IQA, Aesthetics) and overall video quality (FasterVQA, DOVER). This ensures a comprehensive enhancement of the perceptual performance of the model, rather than overfitting to specific metrics. The results in A1-2 further support this point.

Q1-2 Limited NR-IQA diversity: Lacks evaluation with other no-reference metrics like MUSIQ, MANIQA, Q-ALIGN.

A1-2 Thank you for the suggestion. We conduct more evaluations and clarify the reasons for only CLIP-IQA.

(1) Evaluation. We evaluate our method (DOVE), and current SOTA approaches (Upscale-A-Video, MGLD-VSR, VEnhancer, and STAR) on the UDM10 dataset using additional NR-IQA metrics: MUSIQ, MANIQA, and Q-Align. The results are as follows:

Our method also achieves strong performance, confirming the absence of bias.

(2) Clarification.

In the paper, we apply a wide range of metrics, including FR-IQA, NR-IQA, and VQA, to ensure comprehensive evaluation. Due to space limitations and the fact that NR-IQA typically measures only single-frame quality rather than full video quality, we primarily report the representative NR-IQA (CLIP-IQA) results.

Thanks again for your suggestion. We will include more no-reference metrics in the revision.

Q1-3 No subjective quality study: Lacks human evaluation (e.g., MOS) to support perceptual quality claims. Over-reliance on pretrained prior: Performance gains may stem largely from CogVideoX, not from method design alone.

A1-3 Thank you for the comment. We address both concerns by conducting a user study and clarifying the role of pretrained priors.

(1) User-Study.

Setup. We conduct a user study involving 20 volunteers. DOVE is compared against four SOTA methods: RealBasicVSR, Upscale-A-Video, MGLD-VSR, and STAR. We randomly select 50 videos from three real-world datasets (RealVSR, MVSR4x, and VideoLQ). Each participant is shown 10 randomly selected samples (from the 50 videos). We show the LR input and recovery results of all methods for LR (shown in a randomly anonymous order). Participants are asked to vote for the best quality result for each sample.

In total, we collect 200 votes. The results are as follows:

Result. DOVE receives 128 votes, accounting for 64% of the total, significantly outperforming other methods. These findings align with the quantitative and qualitative results in the paper.

(2) Pretrained Prior.

1. Leveraging pretrained models is a feature of our method and also a corresponding advantage. Besides, many recent VSR approaches also build upon pretrained backbones.
2. However, to achieve the performance of DOVE, it is not enough to only utilize pretrained models. It depends on our proposed latent-pixel training strategy and the tailored video processing pipeline (i.e., HQ-VSR). Tables 1a and 1c in the main paper highlight the contributions of these designs.
3. Moreover, the non-specific model design feature of our method also ensures generalizability. To validate this, we extend our method (training strategy and dataset) to another pretrained T2V model, WAN2.1-T2V-1.3B [ref1]. The resulting model (denoted as DOVE (WAN2.1)), trained with our approach, also performs well. We show results on UDM10:

[ref1] Wan: Open and Advanced Large-Scale Video Generative Models, 2025.

Q1-4 How does DOVE perform on un-seleted no-reference IQA metrics such as MUSIQ, MANIQA, and Q-ALIGN? Since the current evaluation mainly reports metrics (CLIP-IQA, DOVER, FasterVQA) that were also used in data filtering, it raises concerns about potential bias. To better assess the model’s generalization in perceptual quality, it would be valuable to compare DOVE against other diffusion-based baselines (Upscale-A-Video, MGLD-VSR, VEnhancer, STAR) under these additional NR-IQA metrics. This would help confirm whether DOVE maintains its superiority across diverse perceptual criteria not seen during data construction.

A1-4 Thank you for the comment.

The metrics used in data filtering are general and diverse, which minimizes the risk of selection bias. We provide additional evaluations using other NR-IQA metrics: MUSIQ, MANIQA, and Q-Align.

For details, refer to A1-1 and A1-2.

Q1-5 Is the choice of t = 399 in one-step inference empirically justified? Providing a comparison with other values could help readers better understand the effect of this hyperparameter and assess whether the choice generalizes across datasets or degradation types.

A1-5 Thank you for the suggestion.

(1) Ablation on Different Values of $t$. We conduct an ablation study on the denoising step t, as reported in Tab. 3 of the supplementary material. We show part of the primary results below:

Setting $t=399$ achieves a better balance between fidelity and perceptual quality. These results align with our analysis in Sec. 3.1 in the main paper. More detailed results and analyses are in the supplementary material.

(2) Generalization across Datasets (REDS). We further conduct the same experiment on another training dataset (REDS), with other settings remaining unchanged. The results are as follows:

The results are close to those trained on HQ-VSR, which shows that the choice of $t$ is generalizable.

Q1-6 Will the authors release HQ-VSR, pretrained models, and training code for reproducibility?

A1-6 Yes, we will release the code, HQ-VSR dataset, and pretrained models to support full reproducibility.

Dear Review 1Gww,

Thank you for taking the time to review our paper and for your thoughtful and constructive comments. We have carefully considered your comments and provided detailed responses to your questions:

1. We clarify our selection of NR-IQA metrics and conduct evaluations using more no-reference metrics to demonstrate the perceptual generalization ability of our method.
2. We perform a user study to support the perceptual quality. We also provide a more detailed explanation of using pretrained diffusion priors in our method.
3. We conduct an ablation study on the timestep t to justify our choice.

As the author–reviewer discussion phase is drawing to a close, we would like to confirm whether our responses have addressed your concerns. If you have any questions or suggestions, please let us know. We would be grateful to hear them.

Thank you again for your time, effort, and valuable feedback.

Best regards,

Authors

Dear Review 1Gww,

Thank you for your response. First, we are pleased that you appreciated our response and experiment.

Second, regarding your suggestion to include NR-IQA evaluation on the real-world dataset, we fully agree that this is both meaningful and valuable. We provide results on the VideoLQ dataset below:

Our method performs well across diverse NR-IQA metrics, further supporting its generalization. We will incorporate these results in the revision of the paper to provide a more comprehensive evaluation.

Thank you again for your thoughtful feedback and helpful suggestions.

Best regards,

Authors