QoE Comparison from Real to Synthetic: Cloud Gaming Traffic Generation and QoE Modeling

Cloud gaming research is constrained by the limited availability of open, reproducible, and quality-of-experience (QoE)-labeled traffic traces. Commercial platforms are closed, gameplay behavior is difficult to reproduce, and network conditions affect both video quality and interactivity. This paper presents CGSynth, a QoE-preserving framework that reconstructs packet-level cloud gaming traces while maintaining the principal quality trends observed in real gameplay sessions. We capture sessions from three representative games—Fortnite, Forza Horizon 5, and Mortal Kombat 11—streamed through Xbox Cloud Gaming and replay them under five controlled bandwidth conditions ranging from 2 to 10 Mbps. Packet timestamps are synthesized using a lightweight hybrid 1D-CNN–GRU model, while payloads are regenerated through standards-compliant video encoding and RTP packetization to preserve the uplink–downlink interaction structure. We also introduce an objective cloud gaming QoE model that combines video-quality metrics, including VMAF, SSIM, PSNR, LPIPS, and smoothness, with interactivity metrics covering response time and synchronization. The synthetic traces reproduce the monotonic QoE–bandwidth relationship observed in the real sessions, with average QoE deviations of 5–10% at most operating points and worst-case deviations below 12%. Video-quality differences are typically below 1 dB in PSNR and 10% for the evaluated perceptual metrics, while interactivity discrepancies decrease to below 5% at higher bandwidths. A controlled study with 20 participants further shows agreement between objective QoE and normalized subjective MOS, yielding Spearman correlations of 0.964, 0.804, and 0.964 for Fortnite, Forza Horizon 5, and Mortal Kombat 11, respectively. Comparison with an empirical sampler and NetDiffusion shows that, although the baselines reproduce some marginal packet statistics more accurately, CGSynth additionally generates decodable gameplay video and preserves command–video alignment for end-to-end QoE evaluation. These results support CGSynth as a reproducible framework for controlled cloud gaming experimentation within the evaluated games, platform, and network conditions.