Recurrent Sinusoidal INRs for Efficient High-Fidelity Representation
A shared sinusoidal recurrent block is presented as a way to enrich INR spectra with less compute.
The paper argues that sinusoidal activations create a harmonic line spectrum, and that recurrent unrolling expands the effective spectral support. The proposed model repeatedly refines a latent representation through that shared block. In tests, it beats feed-forward INR baselines on RGB image benchmarks with fewer parameters and fewer optimization steps. The authors also report favorable transfer to super-resolution, NeRF, and SDF tasks. HF Daily Papers' note
The paper argues that sinusoidal activations create a harmonic line spectrum, and that recurrent unrolling expands the effective spectral support. The proposed model repeatedly refines a latent representation through that shared block. In tests, it beats feed-forward INR baselines on RGB image benchmarks with fewer parameters and fewer optimization steps. The authors also report favorable transfer to super-resolution, NeRF, and SDF tasks. HF Daily Papers' note
score 4