Nexels: Neurally-Textured Surfels for Real-Time Novel View Synthesis (lessvrong.com)

Nexels, a novel approach to real-time novel view synthesis, introduces neurally-textured surfels that efficiently represent fine appearances in 3D scenes using significantly fewer primitives. Traditional Gaussian splatting relies on large numbers of colored primitives to depict intricate details, whereas Nexels employ a shared neural field model—specifically Instant-NGP—to capture detailed textures while drastically reducing the required count from millions to just 4,000. This innovation not only enhances rendering efficiency but also maintains high perceptual quality even with fewer primitives, demonstrating superior performance over conventional methods. The significance of Nexels for the AI and machine learning community lies in its potential to streamline 3D rendering processes while providing rich visual content. By utilizing a two-pass rendering strategy that selectively texturizes only the highest-weighted primitives, Nexels optimizes computational resources without sacrificing quality. This novel dataset, which features close-up shots and diverse scene details, further validates the model's capabilities in retaining photometric quality. As a result, Nexels not only simplifies the technical requirements of rendering but also represents a leap forward in the way fine appearance details are synthesized in computer-generated imagery.