TwinVFI-PPO: Uncertainty-Aware Joint Bitrate Adaptation and Frame Interpolation for Neural Video Streaming

Abstract

Adaptive bitrate (ABR) control and client-side video frame interpolation (VFI) are usually optimized as separate modules, even though they compete for the same playback, bandwidth, and compute budgets and jointly determine rendered quality. We present TwinVFI-PPO, an uncertainty-aware framework that selects bitrate and VFI mode as a unified action. A probabilistic network digital twin estimates short-horizon throughput distributions, a motion-confidence VFI branch predicts restoration gain and tail latency, and a feasibility-aware PPO policy suppresses bitrate–mode combinations that cannot satisfy the available presentation budget. The resulting controller optimizes encoded quality, rebuffering, temporal variation, interpolation latency, and uncertainty within a single sequential decision process. On Vimeo90K and UCF101, the proposed VFI branch reaches 36.17/0.986 and 38.29/0.975 PSNR/SSIM at 53 ms for 720p inference. Across FCC+Norway, OBOE, and Live no-loss replays, TwinVFI-PPO obtains average QoElin and QoElog scores of 47.72 and 46.86, outperforming PPO-ABR by 3.29% and 2.72%. Component analysis further shows a 13.1% bitrate increase together with 38.5% lower rebuffering and 53.3% lower smoothness penalties. These results demonstrate the value of coordinated network–compute control for perceptual neural video streaming.