High-resolution image representation using sinusoidal representation networks and a custom activation variant.
Highlights
Reproduced SIREN to fit 1024×1024 images as continuous implicit neural representations.
Designed and analyzed a custom activation function to replace the sinusoid and compared its reconstruction quality and convergence behavior against SIREN.
Ran ablation experiments to validate hypotheses about why sinusoidal activations work so well for implicit image representations.