Compressed Domain Prior-Guided Video Super-Resolution for Cloud Gaming Content

Qizhe Wang, Qian Yin, Zhimeng Huang, Weijia Jiang, Yi Su, Siwei Ma, Jiaqi Zhang

Introduction

Cloud gaming represents a paradigm shift in the gaming industry, where high-fidelity graphics are rendered on powerful remote servers and streamed to end-user devices. To maintain low latency and manage bandwidth, these video streams are often transmitted at lower resolutions and subjected to heavy compression. At the client side, video super-resolution (SR) is employed to restore the content to high resolution (HR). However, traditional SR methods struggle with two primary issues in this context: the amplification of compression artifacts (such as blocking and ringing) and the high computational cost that often prevents real-time execution at the necessary 60+ frames per second (FPS).

To address these challenges, researchers from Peking University and Migu Interactive have introduced CPGSR (Coding Prior-Guided Super-Resolution), a lightweight network designed specifically for cloud gaming. By utilizing metadata from the video compression process—known as coding priors—CPGSR effectively suppresses artifacts while maintaining the efficiency required for consumer-grade hardware.

The CPGSR Architecture

The core innovation of CPGSR lies in its ability to use "hints" from the video encoder to guide the reconstruction process. Unlike standard SR models that only see decoded pixels, CPGSR integrates several key components to leverage compressed domain information:

Compressed Domain Guided Block (CDGB)

The CDGB serves as the primary feature extraction unit. It incorporates coding priors, including the Quantization Parameter (QP), prediction signals, residuals, and partition maps. These priors act as a roadmap, informing the model where the encoder struggled and where actual object boundaries exist, allowing the network to distinguish between true image detail and compression-induced noise.

Pixel-Adaptive Convolution (PAC)

To mitigate blocky artifacts, the model employs Pixel-Adaptive Convolution. By utilizing the video's partition map—the grid showing how the encoder divided the frame into blocks—the PAC adjusts convolution weights spatially. This allows the network to apply different processing strategies to flat textures versus sharp edges, ensuring that object boundaries remain crisp while compression artifacts are smoothed out.

Structural Re-parameterization (Repconv)

To meet the ultra-low latency requirements of cloud gaming, the authors utilize structural re-parameterization. During the training phase, the model uses complex "Repconv" blocks with multiple branches to capture intricate features. Before deployment (inference), these branches are mathematically collapsed into a single, efficient $3\times3$ convolution. This technique provides the representational power of a deep network with the execution speed of a much shallower one.

Frequency-Aware Optimization

Standard loss functions like Mean Squared Error (MSE) often fail to capture the high-frequency details that the human eye perceives as sharpness. To combat this, the researchers developed the Partitioned Focal Frequency Loss (PFFL). Inspired by the block-based nature of video coding, PFFL forces the model to focus on recovering high-frequency information within $32\times32$ coding units. This targeted optimization helps restore the fine details typically lost during the quantization process of video encoding.

Experimental Results and Dataset

Because existing SR datasets do not typically include the metadata required for compressed-domain research, the authors created a new compressed cloud gaming dataset. This dataset features 10 different game sequences at 1080p, downsampled and then compressed using the Versatile Video Coding (VVC) standard.

CPGSR demonstrated superior performance compared to nine other state-of-the-art lightweight SR models:

  • Efficiency: The model is remarkably lean, containing only 174K parameters. For comparison, the IMDN model requires 694K parameters (4x more) to achieve similar quality.
  • Image Quality: CPGSR achieved a PSNR-Y of 35.02 dB, outperforming models like RLFN and RFDN. It showed particularly strong results in chromaticity restoration (PSNR-U and PSNR-V).
  • Real-time Performance: The architecture supports inference speeds exceeding 60Hz, making it a viable solution for real-world cloud gaming deployments on consumer devices.

Conclusion

CPGSR demonstrates that metadata from the video compression process is a powerful tool for enhancing super-resolution performance. By integrating coding priors and utilizing structural re-parameterization, the model achieves a high-quality upscaling effect while remaining lightweight enough for real-time gaming. While currently optimized for the VVC standard, the principles behind CPGSR offer a promising path forward for utilizing compressed-domain information in various video processing tasks.

Summary updated

BibTeX
@inproceedings{wang2025compressed,
  author = {Qizhe Wang and Qian Yin and Zhimeng Huang and Weijia Jiang and Yi Su and Siwei Ma and Jiaqi Zhang},
  title = {{Compressed Domain Prior-Guided Video Super-Resolution for Cloud Gaming Content}},
  booktitle = {DCC},
  year = {2025},
  doi = {10.1109/dcc62719.2025.00016},
}
Figure