AI-Assisted Super-Resolution and Resource Optimization in Cloud Gaming

Abstract

Cloud gaming (CG) has transformed the gaming industry by enabling users to remotely access resource-intensive games on lightweight devices. However, delivering a seamless gaming experience remains challenging due to stringent bandwidth requirements, at least 25 Mbps for 1080p streaming, and strict latency tolerances that depend on game genre, approximately 80 ms for fast-paced first-person games and 100-150 ms for slower third-person and story-driven titles. This thesis addresses these challenges through the systematic development of deep learning-based super-resolution (SR) solutions integrated within the CG pipeline to reduce bandwidth consumption while preserving visual quality and interactivity.

We present four interconnected contributions. First, we propose GameSR, a lightweight neural super-resolution model that operates directly on encoded game frames without requiring game engine integration or source code access. By combining reparameterized convolutional blocks with a lightweight ConvLSTM for temporal learning, GameSR achieves real-time performance of up to 240 FPS on a GPU-accelerated client at comparable perceptual quality to native streaming.

Second, we develop ARCADE, an adaptive cloud gaming framework that jointly optimizes rendering resolution, encoding bitrate, and client-side super-resolution through offline reinforcement learning. By rendering at a lower resolution when scene complexity permits, ARCADE saves server-side resources on two fronts simultaneously: it reduces the GPU cost of rendering the frame and the cost of compressing (encoding) it, while the lower bitrate cuts transmission bandwidth. This yields reductions in server CPU and GPU utilization of up to 62% and 41% respectively, alongside up to 50% bandwidth savings compared to standard WebRTC streaming at equivalent or higher visual quality.

Third, we extend our approach to immersive virtual reality cloud gaming. We demonstrate that transmitting one color view and one monochrome view reduces VR streaming bandwidth by up to 56% (iSR), and address the resulting sequential pipeline bottleneck by developing GameSRVR, a unified model that jointly performs stereo-aware colorization and super-resolution in a single forward pass. GameSRVR reduces inference from 73.61 ms to approximately 11 ms per stereo pair at 4× scaling, a 6.7× speedup. Deployed on a cloud gaming testbed, GameSRVR achieves over 33% bandwidth savings and an end-to-end latency of approximately 20 ms, meeting the VR motion-to-photon budget.

Fourth, to enable reproducible AI-driven cloud gaming research, we introduce GameLab, an open-source, AI-enabled cloud gaming testbed built on WebRTC. GameLab provides programmable server-side and client-side hook points for integrating machine learning modules, along with a novel QR-based frame identity mechanism enabling reliable full-reference quality evaluation under real network conditions.

Extensive objective evaluations on popular games including Counter-Strike 2, Overwatch 2, Team Fortress 2, and FIFA 24, as well as VR titles Beat Saber, Pavlov VR, and VRChat, alongside subjective user studies, validate that our solutions deliver high perceptual quality while meeting real-time constraints. This thesis demonstrates that effective cloud-gaming neural upsampling can be achieved without proprietary game engine integration, offering practical solutions deployable across both legacy and modern game titles.