Mechanistic electrophysiology simulations provide detailed insights into arrhythmia mechanisms but are hindered by high computational cost. Recent emergence of reaction-eikonal models rely on efficient surrogates for the diffusion current to maintain scalability, although existing approximations struggle to generalize across heterogeneous conduction properties and complex arrhythmic dynamics. This paper presents adaINR -- a novel implicit neural representation (INR) for the diffusion current that is adaptive to local conduction characteristics. Trained and tested on data from a simulated reentry scenario in cardiac tissue, we demonstrated the ability of adaINR to efficiently capture the key features of diverse diffusion current morphologies - including planar, curved, and colliding wavefronts - with an average mean squared error of 33.8 (\muA/cm^2)^2. It has the potential to facilitate faster and more scalable simulations of complex arrhythmias.