2. Image Fitting

Reproducibility details. We fit a 300×300 gray image for 100 epochs using baseline SIREN with ω₀=3000 and SIREN² with four hidden layers of 256 features each. Optimization uses Adam with an initial learning rate 1e-4 that decays by 0.2% every 20 epochs. It took 0.8 seconds to train each image on an NVIDIA L40s GPU (48GB).

For a fixed target signal, the PSNR gain from our initialization scheme (used in SIREN²) increases with parameter count; that's why, with the same network, grayscale training yields a larger percentage PSNR improvement than RGB.

More experiments in the paper.