NNiT: Width-Agnostic Neural Network Generation with Structurally Aligned Weight Spaces

The paper introduces Neural Network Diffusion Transformers (NNiTs), which enable width-agnostic generation of neural network weights by tokenizing weight matrices into patches and utilizing Graph HyperNetworks (GHNs) with a CNN decoder for structural alignment. This method allows for the generation of fully functional Multilayer Perceptrons (MLPs) across various architectures, achieving over 85% success on unseen architecture topologies in ManiSkill3 robotics tasks, while traditional approaches struggle with generalization. This advancement is significant for practitioners as it facilitates the creation of adaptable neural network architectures that can effectively generalize across diverse applications.