This work was conducted at CMLab, Chung-Ang University (advisor Prof. Jihyong Oh). The first author, Sukhun Ko, is an M.S. student at the Graduate School of Advanced Imaging Science, researching INRs, diffusion-based generation, and 3D reconstruction. Co-authors are Seokhyun Youn, Dahyeon Kye (CMLab), Prof. Chanho Eom (Perceptual AI Lab), and Kyle Min (Oracle), and the corresponding author is Prof. Jihyong Oh. The team studies visual signal representation based on frequency analysis.

Sukhun Ko notes that Implicit Neural Representations (INRs) lack frequency selectivity and spatial localization, causing spectral bias. Sukhun Ko proposes FLAIR, whose Band-Localized Activation (BLA) performs frequency selection and spatial localization under the time-frequency uncertainty principle, and Wavelet-Energy-Guided Encoding (WEGE) provides explicit frequency guidance via wavelet energy with only 0.1K parameters. FLAIR outperforms prior INRs on 2D image tasks, 3D reconstruction, and novel view synthesis.