TAQ: Static-Deployable Temporal-Aware Quantization for Real-World Video Super-Resolution

Real-world video super-resolution (VSR) faces in-the-wilddegradations whose artifacts accumulate over time, yet edge deploymentrequires static execution, and quantization parameters must be fixed af-ter compilation for low-overhead inference. We propose Temporal-AwareQuantization (TAQ), a novel static post-training quantization frame-work that uses video structure only in offline calibration. TAQ calibratessequence-specific activation bounds, refines them with a temporal con-sistency objective we propose that aligns inter-frame changes betweenfloating-point and quantized outputs without weight retraining, and en-sembles the refined bounds into one deployable set of static param-eters. Across REDS, SPMCS, UDM10, and VideoLQ, TAQ improvesperceptual and temporal quality, reducing LPIPS by up to 0.0334 andTLPIPS by up to 4.7344 over Static PTQ baselines under identical cal-ibration. On NVIDIA Jetson Orin Nano, TAQ enables TensorRT INT8with up to 3.56× speedup (1.15 fps), while a dynamic INT8 baseline isslower than FP32 (0.79×, 0.266 fps vs. 0.325 fps). Code is available athttps://github.com/imaboybut/TAQ.