HyperNeRF: A Higher-Dimensional Representation for Topologically Varying Neural Radiance Fields

Park et al. extend their earlier Nerfies (Park et al., 2020) deformable radiance-field model to scenes with topological change — an opening mouth, a fist that unclenches, a knife slicing fruit. The Nerfies-style deformation field maps each observed point into a single canonical template; this works for elastic deformation but fails when the topology itself changes, because a continuous deformation field cannot represent a discontinuity (an opening or closing of a connected component).

HyperNeRF resolves the discontinuity by lifting the canonical template into a higher-dimensional ambient space. Each input frame is associated with an ambient coordinate; the radiance field is then defined over the joint (3D position, ambient code) space, and the per-frame appearance is recovered by slicing the higher-dimensional field at the frame's ambient location. The construction is inspired by level-set methods in classical geometry, where evolving surfaces are modeled as slices of a higher- dimensional implicit function.

The deformation field is retained from Nerfies — the network first warps observed points into the canonical 3D coordinates, then performs the ambient slice — and the elastic regularization and coarse-to-fine positional encoding of that earlier work are reused. On a benchmark of casually captured topology-changing clips, HyperNeRF reduces LPIPS error by 4.1% for moment-interpolation and 8.6% for novel-view synthesis compared to Nerfies, and renders cleanly across topology transitions.

The higher-dimensional canonical-space construction has been adopted by subsequent monocular dynamic methods that struggle with topology, and the Nerfies / HyperNeRF pair remains the canonical pre-Gaussian reference for deformable radiance fields. Both works appear together in the dynamic-3D portion of the corpus mapped here.