Reproduce “Nexels: Neurally-Textured Surfels for Real-Time Novel View Synthesis with Sparse Geometries” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Victor Rong, Jan Held, Victor Chu, Daniel Rebain, Marc Van Droogenbroeck, Kiriakos N. Kutulakos, Andrea Tagliasacchi, David B. Lindell
Nexels separates scene geometry and appearance using oriented surfels and a shared neural field for detailed view-dependent texture. It limits field queries to the K highest-weight surfels per pixel. On five Mip-NeRF360 outdoor scenes, 400,000 Nexels achieve LPIPS 0.230, versus 0.234 for 3DGS with 3.9 million Gaussians, showing comparable measured perceptual quality with fewer primitives in this setting.
Key insights
- Decoupling explicit geometric primitives from fine-grained appearance modeling significantly reduces scene representation size without sacrificing visual fidelity.
- Introducing a differentiable `gamma` parameter allows surfels to effectively model both soft Gaussian blends and sharp quad-like surface boundaries.
- A two-pass rendering strategy that selectively applies neural textures only to a fixed, small number of highly-contributing primitives per pixel enables real-time performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.