RTGSR: Real-Time Game Content Super-Resolution via Compressed-Domain Coding Priors

The rapid growth of cloud gaming and game streaming has led to a substantial increase in the volume of game content data. To ensure real-time delivery of cloud game content, a common strategy is to downsample and compress the game content before transmission, reducing both data size and bandwidth requirements. However, this approach presents considerable obstacles for super-resolution (SR) networks at the receiver side. In particular, the degraded quality of compressed video streams, combined with the stringent demand for real-time processing, poses major challenges for practical SR applications. In this paper, we propose a novel real-time super-resolution framework that works directly in the compressed domain by exploiting coding-domain priors. Specifically, we propose an extremely lightweight U-Net architecture that leverages prediction maps and residuals as its primary guidance signals. Furthermore, we incorporate the partition map into a Pixel Adaptive Convolution (PAC) module, allowing the convolution kernels to adapt to different regions in the decoded frame. The resulting deep features are then fused with those from the U-Net backbone through an attention block. Finally, we present an enhanced re-parameterization block designed to better model edge features, leading to notable gains in both the objective metrics and subjective visual quality of the reconstructions. Extensive experiments demonstrate that the proposed method consistently outperforms existing real-time approaches on compressed game video content, achieving superior performance in both quality and efficiency.