Recurrent Sinusoidal INRs for Efficient High-Fidelity Representation

Researchers introduced a recurrent sinusoidal implicit neural representation that iteratively refines latent features with a shared block.

The authors argue sinusoidal activations build harmonic spectral enrichment across unrolling, improving representation capacity.

The method outperforms feed-forward and other recurrent baselines on images, super-resolution, NeRF, and SDF tasks.

It also achieves strong fidelity with fewer parameters and fewer optimization steps, making it more efficient for compact high-quality representations.