Outlier-Aware Post-Training Quantization for Image Super-Resolution

Post-training quantization for SR fails where the outliers are

Quantization is essential for deploying super-resolution networks on edge devices. Quantization-aware training (QAT) works well but needs ground truth and hours of retraining; post-training quantization (PTQ) needs neither, yet existing PTQ methods for SR lag far behind — because they overlook what happens to activation outliers.

Our empirical analysis reveals that activation outliers are strongly correlated with image colour information: removing them causes visible colour shifts and severe performance degradation. But retaining them naively consumes most of the available bit-width, starving the normal activations. We resolve this trade-off with a dual-region quantization strategy that partitions activations into an outlier region and a dense region and quantizes each independently. We further observe that different network layers vary widely in quantization sensitivity, and introduce a sensitivity-aware finetuning loss that focuses the optimization on the layers that matter. Extensive experiments show that our method outperforms existing PTQ approaches across SR networks and datasets, while matching QAT in most scenarios at a 75× speedup.

BibTeX

@inproceedings{wang2025outlier,
  title     = {Outlier-Aware Post-Training Quantization for Image Super-Resolution},
  author    = {Wang, Hailing and Lu, Jianglin and Zhang, Yitian and Fu, Yun},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  year      = {2025}
}

Acknowledgements. Our implementation builds on AdaBM and EDSR-PyTorch, and the piecewise quantizer primitives follow PWLQ. We thank the authors for releasing their code.