MLP Splatting: 3D scenes from hundreds of MLP primitives instead of...

MLP Splatting: 3D scenes from hundreds of MLP primitives instead of millions of Gaussians

Cambridge researchers introduced MLP Splatting (Object-Centric Neural Fields) — a 3D scene representation accepted at ECCV 2026, where each primitive is a compact MLP and, when trained only on RGB, aligns to objects and their parts without segmentation, matching Feature-3DGS in quality with 63 times fewer primitives and 6 times less memory.

Compiled by Sergey KostenchukPublished 2026-10-16Updated 2026-10-17

2026-10-16 Research

🤖 MLP Splatting: 3D scenes from hundreds of MLP primitives instead of millions of Gaussians

The Object-Centric Neural Fields method from Cambridge (Shinjeong Kim, Andrew J. Davison and colleagues) was accepted at ECCV 2026, preprint arXiv:2606.03877. Each scene primitive is a compact MLP with a Gaussian, predicting color and opacity.

🌍 When trained only on RGB, the primitives themselves settle onto objects and their parts — no segmentation is needed. Compared with Feature-3DGS: 63 times fewer primitives, 6 times less memory, and rendering twice as fast at comparable quality (Replica, ScanNet).

👤 Open-source code under MIT (CUDA, GPU 8.0+, Python 3.10+, torch>=2.10.0): you can train a model on Replica/ScanNet scenes and move or recolor objects by selecting a pair of primitives; LSeg feature distillation adds open-vocabulary segmentation and object removal.

Source 1: http://shinjeongkim.com/mlp-splatting Source 2: https://github.com/kxhit/mlp-splatting

Sources