Slice: A Selective Local Inference Framework with Codec Exploitation for Accelerating Video Super-Resolution

Video now constitutes the majority of network traffic, and users expect lag-free, high-quality playback. Since network bandwidth often cannot sustain high-quality streams, client-side Super-Resolution has become an attractive option. In particular, neural network-based Super-Resolution offers strong perceptual gains by recovering fine detail. However, running perframe inference is costly on edge hardware and often pushes latency and power budgets beyond their limits. When server assistance is available, server-orchestrated Super-Resolution pipelines that decide the upscaling strategy in the server side can effectively reduce per-frame cost, but settings such as live streaming and on-device playback often preclude such support, highlighting the need for a client-only alternative. Since the standard bitstream provides the codec metadata to reconstruct the bits into pixel-level frames, we find that these codec metadata can act as a cue to accelerate the client-only Super-Resolution pipeline. Specifically, we observed that motion vectors and residuals can be used to locate regions where Super-Resolution yields the largest quality gains and to reveal unchanged areas that can be reused from the previous high-resolution frame. This enables selective patch-level inference on informative regions while reusing stable content and interpolating the rest, which keeps per-frame latency within budget with minimal quality loss. Based on these observations, we present Slice, a framework that applies different upscaling strategies to patches within a frame according to their codec-derived characteristics. Slice runs entirely on the client using only client-side information and can further exploit energy-efficient hardware decoder by leveraging codec metadata available in the standard bitstream. Across real-world videos, Slice delivers 2.72× speedup and reduces energy consumption by 62.57% compared to per-frame inference.