Smooth Real-time Rendering via Implicit Nested Neighborhoods

Vinícius Da Silva, Tiago Novello, Guilherme Schardong, Luiz Schirmer, HELIO CORTES VIEIRA LOPES, Nuno Gonçalves, Alberto Raposo, Luiz Velho

Keywords: Implicit Neural Representations, Neural Signed Distance Functions, Neural Rendering, Real-time Rendering, Surfaces and Attributes

TL;DR: We propose an end-to-end Implicit Neural Representation framework to render surfaces in real-time using smooth ($C^{\infty}$) neural signed distance functions (SDFs) endowed with smooth attributes such as normals and textures.

Abstract:

Implicit neural representations (INRs) for surfaces have been mostly used as intermediary representations before triangle mesh extraction. Extracting meshes is not a real-time task and introduces unnecessary discretization to rendering, making it difficult to fully use the smoothness of INRs in applications. Smooth INRs are broadly used for approximating surface \textit{signed distance functions} (SDFs) through an implicit regularization (Eikonal equation) using their available high-order derivatives. Such property also makes it easier to integrate those INRs in pipelines that explore differentiable properties of the underlying surface. The current real-time state-of-the-art approach uses grid-based data-structures that introduce discretization, resulting in a non-smooth representation.

We propose an end-to-end smooth ($C^{\infty}$) INR framework to represent and render surfaces in real-time using neural SDFs endowed with smooth attributes such as normals and textures. Our approach leverages from a novel localized SDF training based on nested neighborhoods, a multiscale surface representation, and residual training. The framework does not depend on spatial data-structures, nor surface extraction. We show that our representation renders detailed smooth surfaces in real-time while the previous works can only render coarse non-smooth surfaces. We also present applications of our representation, including integration with a pipeline for dynamic surfaces and a way to improve performance of surface extraction via marching cubes.

Primary Area: applications to computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 11947