HarmoQ: Harmonized Post-Training Quantization for High-Fidelity Image Super-Resolution

Modern image super-resolution (SR) networks match photographic detail but remain impractical for memory- and latency-constrained devices. Post-training quantization (PTQ) compresses such models without full retraining; yet standard PTQ treats weight and activation quantizers independently and therefore ignores their coupled behaviour in the frequency domain. A Fourier analysis reveals a consistent asymmetry: quantizing weights blurs colour and shape at low frequencies, while quantizing activations erodes edges and textures at high frequencies. Because these distortions interact multiplicatively rather than add linearly, separate optimisation leaves considerable perceptual quality unrealised. We introduce HarmoQ, a harmonised PTQ framework that accounts for this coupling. HarmoQ first applies a closed-form spectral residual projection, editing weights once to cancel high-frequency artefacts caused by activation clipping. It then enforces an equal-error scaling law that analytically balances mean-squared error between weights and activations, followed by an adaptive boundary refinement routine that fine-tunes clipping thresholds while periodically restoring that balance.