3D Affine Transformations on Neural Fields

Creating neural network architectures that use weights and biases of another neural network as an input is a new and promising field. Possible applications are prediction of input network generalization, classifying implicit neural representations, or style editing.

In this work, we used this approach to apply affine transformations to SIRENs, representing 2D and 3D shapes, using a HyperNetwork that changes the MLP weights with given transformation.