GameSR: Real-Time Super-Resolution for Interactive Gaming

ICLR 2026 Conference Submission20748 Authors

Keywords: Super-Resolution, Neural Upsampling, Gaming

Abstract: High-resolution gaming demands significant computational resources, with challenges further amplified by bandwidth and latency constraints in cloud gaming. Existing upscalers, such as NVIDIA DLSS and AMD FSR, reduce rendering costs but require engine integration, making them unavailable for most titles, especially those released before the introduction of upscalers. We present \textbf{GameSR}, a lightweight, engine-independent super-resolution model that operates directly on encoded game frames. The architecture of GameSR combines reparameterized convolutional blocks, PixelUnshuffle, and a lightweight ConvLSTM to deliver real-time upscaling with high perceptual quality. Extensive objective and subjective evaluations on popular games, such as \textit{Counter-Strike 2}, \textit{Overwatch 2}, and \textit{Team Fortress 2}, show that \system{} reduces cloud gaming bandwidth usage by 30--60\% while meeting target perceptual qualities, achieves real-time performance of up to 240~FPS, substantially outperforms existing super-resolution models in the literature, and reaches near-parity with DLSS and FSR \textit{without} accessing rendering engine data structures or modifying game source code, making \system{} a practical solution for upscaling both modern and legacy games with no additional development effort.

Primary Area: generative models

Submission Number: 20748