Abstract
Universal-filter-based models provide an efficient and interpretable framework for image color correction by predicting parameters of structured correction operators. However, practical deployment still benefits from additional reductions in computational cost and memory footprint, especially for ondevice and real-time pipelines where color stability is perceptually critical. In this work, we study pruning for a universal-filter color correction model and benchmark representative unstructured and structured pruning baselines, including channel pruning and latency-aware structured pruning. We further propose a configuration-agnostic customized sparsity allocation strategy that adapts pruning decisions across predictor components to better preserve quality under compression. Experiments on the MIT-Adobe FiveK benchmark show that unstructured pruning yields limited wall-clock gains despite parameter reduction, whereas structured strategies provide meaningful acceleration. At$\mathbf{5 0 \%}$sparsity, the proposed method reduces latency from 8.5 ms to 4.4 ms and FLOPs from 85.4 M to$\mathbf{3 6. 3 M}$while maintaining high quality (PSNR = 24.06 dB, SSIM$=0.916$). At 60 % sparsity, it further reduces latency to$\mathbf{4. 0 ~ m s}$with stable performance (PSNR$=23.89 ~\text{dB}$, SSIM$=0.910)$. Overall, the proposed strategy achieves a more favorable quality-efficiency trade-off than baseline pruning approaches, improving the deployability of universal-filter color correction models.