High-Frequency First: A Two-Stage Approach for Improving Image INR

Abstract

Implicit Neural Representations (INRs) have emerged as a powerful alternative to traditional pixel-based formats by modeling images as continuous functions over spatial coordinates. A key challenge, however, lies in the spectral bias of neural networks, which tend to favor low-frequency components while struggling to capture high-frequency (HF) details such as sharp edges and fine textures. While prior approaches have addressed this limitation through architectural modifications or specialized activation functions, the potential of complementary training strategies remains relatively underexplored. In this work, we propose an orthogonal direction by directly guiding the training process. Specifically, we introduce a two-stage training strategy where a neighbor-aware soft mask adaptively assigns higher weights to pixels with strong local variations, encouraging early focus on fine details. The model then transitions to full-image training. Experimental results show that our approach is complementary to existing INR methods and consistently improves reconstruction quality. By assigning frequency-aware importance to pixels in image INR, our work offers an effective avenue to mitigate the spectral bias problem.