acceptodds
Under review as a conference paper at ICLR 2027

ScopeVSR: A Broad-Coverage Training Dataset for Real-World Video Super-Resolution

Abstract

Diffusion models have shown superior performance in video super-resolution (VSR). However, restoring diverse real-world content remains challenging due to the lack of balanced semantic coverage in existing training data. In this paper, we introduce a structured curation pipeline for open-domain VSR, SCOPE. In this pipeline, we combine task-oriented video selection with hierarchical semantic rebalancing. Applying this pipeline to a large-scale collection of data sources, we construct the SCOPE-VSR dataset for training and SCOPE-test for testing. SCOPE-VSR has approximately 27k videos with a balanced and diverse coverage of various real-world scenarios and subjects. In addition, we propose GraCo-VSR, a novel framework for one-step video restoration. GraCo-VSR adopts a reliability- aware design with explicit supervision to reduce degradation-induced distortions in motion and texture guidance. This design enables the model to better leverage the diverse training content in SCOPE-VSR, thereby supporting generalization across varied scenarios. Experiments demonstrate that our model exhibits strong performance on both perceptual fidelity and visual quality.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Video Super-Resolution, Video Dataset

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_4oa7OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SCOPE, a curation pipeline that filters video by technical quality, texture and motion, then rebalances scene-subject categories, producing 27,351 training clips from about 0.6 million candidates. It also introduces GraCo-VSR, a one-step restorer with a distilled motion encoder and GT-supervised texture confidence. The paper reports improved distributional balance, benchmark comparisons, a data-pipeline ablation on a 0.1M OpenVid-HD subset, a model-component ablation, and a DOVE fine-tuning comparison between SCOPE-VSR and HQ-VSR.

Critical strengths
  • SCOPE has a reproducible sequence of quality, foreground-motion, and hierarchical scene/subject selection steps. Its top three predicted topic categories account for 31.0% of clips, versus 49.4% for HQ-VSR and 46.5% for YouHQ under the same InternVideo2 labeling procedure.
  • The paper attempts to separate data and model contributions: curation-stage ablations on UDM10, CLDC/RATE ablations with the same backbone and training data, and a DOVE fine-tuning comparison on SCOPE-VSR versus HQ-VSR.
Critical weaknesses
  • SCOPE-test comprises only 40 videos taken from the same candidate collection and curation pipeline. They were left out through category-level rebalancing, and the LR degradation matches training (Sec. 4.1). The paper does not document source-level or near-duplicate separation from training. This test therefore gives limited evidence of transfer to genuinely new sources or degradations.
  • The broad-performance claim is stronger than the tables warrant. On RealVSR, GraCo-VSR has 19.86 PSNR versus 22.16 for RealViFormer and 20.81 for PS-SR, on REDS30 its DISTS is 0.1969 versus 0.1600 for FlashVSR. In the data-only DOVE comparison, SCOPE-VSR improves perceptual measures but lowers PSNR on both YouHQ40 and REDS30. Per-category gains and uncertainty are absent, so the contribution of semantic rebalancing to rare categories remains hard to judge.
Questions for authors
  1. How were source videos, near duplicates, and shared footage excluded between SCOPE-VSR and SCOPE-test and between SCOPE-VSR and the external benchmarks?
  2. Can the authors report per-category test results and confidence intervals for the balance stage, with dataset size and training steps held fixed, and clarify data licenses and release terms for all five source collections?
Reviewer_gTsCOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses two major challenges in open-domain video super-resolution (VSR): imbalanced semantic distributions in training data and unreliable motion and texture guidance caused by degraded inputs. To this end, the authors propose the SCOPE data curation pipeline and the GraCo-VSR restoration framework. SCOPE constructs the SCOPE-VSR dataset, comprising 27,351 videos, through perceptual quality screening, foreground-sensitive motion mining, and hierarchical semantic rebalancing. Built upon the Wan2.1 video diffusion model, GraCo-VSR introduces the CLDC module to extract clip-level motion information and the RATE module to predict texture confidence, thereby suppressing the influence of degradation-induced noise and enabling more reliable one-step video super-resolution.

Critical strengths
  1. Targeted Dataset Construction. The paper proposes a systematic data curation pipeline that combines visual quality screening, foreground-sensitive motion mining, and semantic rebalancing to construct SCOPE-VSR, a dataset containing 27,351 videos. The top three most frequent topics account for 31.0% of the dataset, lower than the corresponding proportions in HQ-VSR and YouHQ, indicating more balanced semantic coverage.
  2. Problem-Oriented Model Design. To address motion estimation errors and texture artifacts caused by degraded inputs, the paper introduces CLDC and RATE to model clip-level motion information and texture reliability, respectively, reducing the impact of unreliable guidance on video restoration.
  3. Ablation Studies Support the Effectiveness of the Main Components. The complete data curation pipeline improves several perceptual metrics, while combining CLDC and RATE further reduces LPIPS and improves DOVER, providing quantitative evidence for the effectiveness of the proposed data curation strategy and model components.
  4. Comprehensive Experimental Evaluation. The paper evaluates the proposed method using multiple metrics on synthetic and real-world video datasets. Furthermore, experiments training DOVE with different training datasets demonstrate the cross-model value of SCOPE-VSR, providing additional evidence for the restoration performance of the proposed method and the generalizability of the dataset.
Critical weaknesses
  1. The paper constructs a relatively large SCOPE-VSR dataset, but provides insufficient details about the data sampling strategy, dataset coverage during training, and actual training time. The authors are encouraged to clarify how video samples are utilized during training and report the associated training costs, enabling a better assessment of the trade-off between the performance gains brought by the large-scale dataset and the additional computational overhead.
  2. The visual consistency and readability of some figures could be improved. For example, Figures 3 and 5 differ in their color schemes, annotations, and overall layouts, while some module labels in Figure 5 are densely arranged. The authors are encouraged to adopt consistent fonts, font sizes, and color schemes, and to refine the layout of the model architecture diagram to improve readability.
  3. Lack of evaluation on real-world videos with unknown degradations and no ground-truth references. The paper primarily evaluates the proposed method quantitatively on test datasets with ground-truth references. However, its performance has not been validated on real-world low-quality videos with complex and unknown degradations and no corresponding ground-truth references, such as VideoLQ. The authors are encouraged to provide additional qualitative comparisons and no-reference quality assessments on such datasets to further demonstrate the generalization capability of the proposed method in real-world scenarios.
Questions for authors
  1. Why does SCOPE use fixed thresholds for video selection? Do these thresholds affect the diversity and coverage of scenes in the dataset?
  2. Why does RATE construct confidence supervision using GT and LR texture information? Can this mechanism effectively distinguish genuine textures from degradation artifacts?
  3. Why does CLDC aggregate optical flow into a clip-level global representation rather than preserving dense motion information? How does this design handle complex local motion and occlusions?
Reviewer_8Rb2OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies open-domain video super-resolution from both data and model perspectives. SCOPE is a structured curation pipeline that filters videos by quality, motion, foreground texture, and semantic distribution. Applied to five public collections, it produces SCOPE-VSR with 27,351 videos and a 40-video SCOPE-test set. The paper also proposes GraCo-VSR, a one-step restoration model based on Wan2.1-1.3B. CLDC provides clip-level motion guidance distilled from optical flow, while RATE predicts texture features and their reliability from degraded inputs. Experiments cover four established benchmarks and SCOPE-test, together with data, module, efficiency, and qualitative analyses.

Critical strengths
  1. The paper addresses an important limitation of VSR training data. SCOPE combines quality filtering, motion and texture selection, and hierarchical semantic rebalancing in a coherent pipeline. The staged data ablation and the additional DOVE experiment provide useful evidence that the curated data can benefit restoration.
  2. GraCo-VSR is well motivated. CLDC and RATE target unreliable motion and texture guidance from degraded videos, and the baseline, +CLDC, +RATE, and combined results show that both modules contribute.
Critical weaknesses
  1. The evidence for the semantic-rebalancing claim is still limited. Figure 4 compares SCOPE-VSR mainly with HQ-VSR and YouHQ, but does not include recent large-scale VSR training sets or the five source pools under the same taxonomy. A broader comparison is needed to show how severe the imbalance is and how much SCOPE changes it.
  2. The ablation of RATE and CLDC is useful but coarse. Table 3 establishes that the two modules help, yet it does not isolate why they help. RATE would benefit from controls for confidence prediction and frequency supervision, while CLDC would benefit from comparisons that separate flow distillation, clip-level aggregation, and multi-scale guidance.
Questions for authors
  1. Could the authors compare the semantic distributions of SCOPE-VSR, recent large-scale VSR training sets, and the five source collections using the same classifier and taxonomy? Please report quantitative measures such as entropy, effective category count, or imbalance ratio, and include a small human audit of the Qwen3-VL scene and subject labels.
  2. Could the authors report category-conditioned VSR results for head, medium-frequency, tail, and worst-performing categories, both before and after semantic rebalancing? This would show whether the balancing step improves the cases it is designed to address.
  3. Could the authors provide more targeted controls for RATE and CLDC?

Comments on the paper

1 comment
Area_Chair_hiJ5OpenReview ↗

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.