Implicit neural fields, typically encoded by a multilayer perceptron (MLP) that maps coordinates (e.g., xyz) to signals (e.g., signed distances), have shown great promise as a compact, high-fidelity representation. . However, the lack of a regular and explicit grid structure also makes it difficult to apply generative modeling directly to implicit neural fields to synthesize new data. To this end, we propose HyperDiffusion, a novel approach for unconditional generative modeling of implicit neural fields. HyperDiffusion operates directly on MLP weights and generates new implicit neural fields encoded by synthesized MLP parameters. Specifically, a collection of MLPs is first optimized to faithfully represent individual data samples. Subsequently, a diffusion process is trained on this MLP weight space to model the underlying distribution of implicit neural fields. HyperDiffusion enables diffusion modeling through implicit, compact, yet high-fidelity representation of complex signals into 3D shapes and 4D mesh animations within a single unified framework.
