INRR: Plug-and-Play and Self-Learned Regularization for Implicit Neural Representations

Implicit Neural Representation (INR) has emerged as a powerful framework for representing continuous functions. However, it still suffers from two limitations: poor generalization to unobserved pixels and high sensitivity to hyper-parameter settings. These issues are particularly pronounced in severely ill-posed recovery tasks. To address these limitations, we propose Implicit Neural Representation Regularizer (INRR), which incorporates two lightweight INRs to adaptively capture non-local self-similarities across the entire data, effectively acting as a self-learnable form of Dirichlet Energy. This allows the main INR model to better capture intrinsic data structures and become more robust to hyperparameter choices. Extensive experiments demonstrate that INRR consistently improves the performance of INR across various inverse tasks, outperforming conventional regularizers. Furthermore, we show that INRR can be readily extended to other signal representation frameworks-such as Deep Matrix Factorization and Deep Image Prior-and applied to diverse data types beyond images, including spatial transcriptomics and recommendation systems. We also present INRR+, an enhanced variant motivated by the twicing algorithm, which further strengthens its representational capability. The proposed INRR offers a generic, plug-and-play regularization strategy that enhances the stability, generalization, and applicability of INR-based models.