HyperDiffusion: Generating Implicit Neural Fields with Weight-Space Diffusion

Implicit neural fields, sometimes encoded by a multilayer perceptron (MLP) that maps from coordinates (e.g., xyz) to alerts (e.g., signed distances), have proven outstanding promise as a high-fidelity and compact illustration. Nonetheless, the shortage of an everyday and specific grid construction additionally makes it difficult to use generative modeling instantly on implicit neural fields with the intention to synthesize new information. To this finish, we suggest HyperDiffusion, a novel method for unconditional generative modeling of implicit neural fields. HyperDiffusion operates instantly on MLP weights and generates new neural implicit fields encoded by synthesized MLP parameters. Particularly, a group of MLPs is first optimized to faithfully signify particular person information samples. Subsequently, a diffusion course of is educated on this MLP weight area to mannequin the underlying distribution of neural implicit fields. HyperDiffusion permits diffusion modeling over a implicit, compact, and but high-fidelity illustration of complicated alerts throughout 3D shapes and 4D mesh animations inside one single unified framework.