GPLMv2: an edge-ready post-training quantization framework for deploying complex underwater image enhancement models

Deploying state-of-the-art underwater image enhancement (UIE) models on resource-constrained marine platforms remains a critical challenge due to their high computational complexity and memory footprints. This paper presents GPLMv2, a framework-agnostic post-training quantization approach designed to bridge the gap between model sophistication and edge hardware constraints. We propose three key innovations: (1) a non-invasive wrapper-based architecture that preserves the functional integrity of custom UIE layers (e.g., attention modules) while enabling INT8 inference; (2) a theoretical derivation revealing that structured quantization acts as implicit regularization, promoting feature diversity essential for generalizing across diverse water conditions; and (3) an adaptation of the GPTQ algorithm incorporating percentile-based clipping and exponential moving average calibration, specifically optimized for the high dynamic range and long-tailed activation patterns inherent in underwater imagery. Extensive experiments on EUVP, UFO, and UIEB datasets demonstrate that GPLMv2 achieves 4x model compression. Counter-intuitively, the quantized model surpasses its full-precision counterpart in enhancement quality, with SSIM increasing from 0.875 to 0.908 on the UFO dataset. Furthermore, the framework successfully compresses the 3.3M-parameter GPLM model from 12.6MB to 3.2MB, making deployment feasible on commodity edge devices like the Raspberry Pi 5. Our theoretical analysis attributes this performance gain to three mechanisms: implicit denoising via quantization noise, regularization-induced generalization, and enforced feature diversity. All models and code are publicly available at https://github.com/Jinxinshao/GPLMv2 to facilitate reproducible research in underwater robotics.