Relighting in neural field techniques

Develop relighting methods for neural field representations, such as Neural Radiance Fields (NeRF), that achieve photorealistic results comparable to the view-synthesis performance of these models.

Background

The paper surveys prior work on neural field methods (e.g., NeRF) and notes that, while these approaches produce highly photorealistic novel view synthesis, they have not yet achieved comparable success for relighting. This gap motivates the authors to propose 3DPR, which combines a 3D generative prior with an OLAT-based reflectance model to perform physically grounded relighting and novel view synthesis from a single portrait image.

By explicitly synthesizing OLATs and linearly combining them with HDR environment maps, the authors aim to address limitations in prior neural field–based approaches that do not provide accurate or controllable relighting, underscoring the need for general relighting solutions within neural field frameworks.