[CVPR 2026 Highlight] CDA-VSR: Compressed-Domain-Aware Online Video Super-Resolution

CDA-VSR is a compressed-domain-aware framework for online video super-resolution.
It explicitly exploits motion vectors, residual maps, and frame types from the compressed bitstream to improve both reconstruction quality and inference efficiency.

News • Overview • Results • Repository Structure • How to use • Citation

---

News

• [2026/03] Repository initialized.
• [2026/03] README released.
• [2026/04] Pretrained model and dataset released.

---

Overview

Online video super-resolution (VSR) aims to recover high-resolution (HR) frames from low-resolution (LR) inputs under strict latency constraints, where only the current and previous frames are available. Most existing online VSR methods operate only on decoded LR frames and often suffer from expensive motion estimation or limited temporal modeling efficiency.

To address this issue, CDA-VSR explicitly introduces compressed-domain priors into online VSR, including:

• Motion Vectors (MV) for efficient coarse alignment
• Residual Maps for selective temporal fusion
• Frame Types (I/P frames) for adaptive reconstruction

Based on these priors, CDA-VSR contains three key modules:

• MVGDA: Motion-Vector-Guided Deformable Alignment
• RMGF: Residual Map Gated Fusion
• FTAR: Frame-Type-Aware Reconstruction

This design enables CDA-VSR to achieve a favorable trade-off between restoration quality and runtime efficiency.

Results

Quantitative Highlights

• On REDS4, CDA-VSR achieves approximately 90 FPS while maintaining competitive or superior reconstruction quality.

Qualitative Comparison

CDA-VSR reconstructs clearer edges and finer textures than representative online VSR baselines, especially in regions with large motion and compression artifacts.

Repository Structure

CDA-VSR/
├── basicsr/                      # Core framework
│   ├── archs/                    # Network architectures
│   ├── data/                     # Dataset definitions
│   ├── losses/                   # Loss functions
│   ├── metrics/                  # Evaluation metrics
│   ├── models/                   # Model wrappers
│   ├── train.py                  # Training entry
│   └── test.py                   # Testing entry
├── options/
│   ├── train/
│   │   └── train_CDA-VSR.yaml
│   └── test/
│       ├── test_CDA-VSR_REDS4.yaml
│       └── test_CDA-VSR_Inter4K.yaml
├──pretrained_models/
    └── best.pth
├── figures/                      # README figures
├── requirement.txt
├── LICENSE.txt
└── README.md

Dataset and Pretrained Models

Dataset

We provide the processed REDS compressed-domain dataset used in this work.

Dataset	Description	Download
REDS_CRF18_23_28_new	Processed REDS dataset with compressed-domain information, including LR frames, motion vectors, residual maps, and frame-type information under different CRF settings.	Baidu Netdisk

Extraction code: zc8e

After downloading, please organize the dataset as follows:

REDS_CRF18_23_28_new/
├── train/
└── test/

Pretrained Models

pretrained_models/
└── best.pth

How to use

Requirements

• Python 3.8+
• PyTorch 1.10+
• CUDA 11.x

Training

Run training with:

python basicsr/train.py -opt options/train/train_CDA-VSR.yaml

Testing

Test on REDS4

Run training with:

python basicsr/test.py -opt options/test/test_CDA-VSR_REDS4.yaml

Test on REDS4

Run evaluation with:

python basicsr/test.py -opt options/test/test_CDA-VSR_Inter4K.yaml

Citation

If you find this repository useful in your research, please cite:

@article{wang2026cdavsr,
  title={Compressed-Domain-Aware Online Video Super-Resolution},
  author={Wang, Yuhang and Li, Hai and Hou, Shujuan and Dong, Zhetao and Yang, Xiaoyao},
  journal={arXiv preprint arXiv:2603.07694},
  year={2026}
}

Contact

Please leave a issue or contact Yuhang Wang with bitwangyuhang@163.com

License and Acknowledgement

This project is built upon BasicSR and TMP. We sincerely thank the authors for making their code publicly available.