Title: MLP Splatting: Object-Centric Neural Fields

Original: arXiv:2606.03877v1 Announce Type: new

Abstract:

Accurate 3D representations are essential for rendering, comprehending, and interacting with digital scenes. While recent techniques like Neural Radiance Fields and 3D Gaussian Splatting have delivered remarkable results in photorealistic novel-view synthesis, they struggle to isolate specific scene components. These methods typically require supplementary segmentation or grouping procedures to facilitate object-level manipulation, as they do not naturally decompose scenes into distinct primitives.

To address this limitation, we introduce MLP-Splatting, a novel approach that facilitates scene decomposition through a small number of highly expressive light-field primitives, all while maintaining high-quality photorealistic rendering. In our framework, each primitive is represented by an independent, compact Multi-Layer Perceptron (MLP) with localized spatial support, responsible for predicting radiance and opacity. Unlike traditional low-level Gaussian primitives or monolithic global radiance fields, our neural primitives offer superior expressive power while maintaining spatial localization. The rendering process leverages efficient sparse volumetric compositing based on ray-primitive interactions.

Through supervision using only RGB data, our method generates primitives that correspond to local scene regions, frequently aligning with distinct objects or object parts. This characteristic allows for interactive, object-level editing without the need for segmentation masks; users can simply select a few primitives to manipulate. Furthermore, by incorporating optional semantic feature distillation, the system supports open-vocabulary scene interaction and instant open-set segmentation.

Experimental results demonstrate that our approach significantly outperforms state-of-the-art methods in efficiency. Specifically, we achieve a 3$\times$ speedup in rendering and reduce memory consumption to just 1/15$\times$ that of comparable semantic 3D Gaussian Splatting methods.

Project Page: https://shinjeongkim.com/mlp-splatting

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC