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

HarmoQ introduces a unified approach to post-training quantization, which is essential for deploying super-resolution models efficiently. The study highlights the critical interplay between weight and activation quantization, revealing that while weight quantization primarily degrades structural similarity, activation quantization disproportionately affects pixel-level accuracy. By systematically analyzing this coupling, the researchers developed HarmoQ, which employs structural residual calibration, harmonized scale optimization, and adaptive boundary refinement to enhance image fidelity. The framework demonstrated substantial gains, outperforming previous methods by 0.46 dB on the Set5 dataset at 2-bit quantization. Additionally, it offers a 3.2x speedup and 4x memory reduction on A100 GPUs, underscoring its practical implications for high-fidelity image processing.