Compressive Neural Representations of Volumetric Scalar Fields

We present an approach for compressing volumetric scalar fields using implicit neural representations. Our represents a field as learned function, wherein network maps point in the domain to output value. By setting number of weights be smaller than input size, we achieve compressed representations fields, thus framing compression type function approximation. Combined with carefully quantizing weights, show that this yields highly compact outperform state-of-the-art volume approaches. The conceptual simplicity our enables benefits, such support time-varying optimizing preserve spatial gradients, and random-access evaluation. study impact design choices on performance, highlighting how simple architectures are effective broad range volumes.