Most radiance field representations do not know where one object ends and the next begins. A Gaussian splatting scene is a dense cloud of small ellipsoids, and a neural radiance field is a single network spread across the whole room, so pulling out a chair or a lamp means adding a separate segmentation step.
MLP Splatting, from Shinjeong Kim, Yuzhou Cheng, Xin Kong, Paul H. J. Kelly and Andrew J. Davison at Imperial College London, builds the scene from far fewer, far more capable pieces. Each primitive is its own small neural network, confined to a soft, stretched region of space, that predicts color and opacity for the rays passing through it. The pretrained room model the team released is made of 2,398 of them.
The primitives are trained on ordinary photographs alone, with no masks or labels, and still tend to settle onto whole objects or object parts. Selecting a handful of primitives is enough to isolate or edit an object. An optional step bakes features from an image and language model into each primitive, which lets a text query find and segment the matching parts of the scene.
Against Feature-3DGS, a Gaussian method that carries the same kind of semantic features, the authors report about 8 frames per second against about 3 and 32 MB of memory against 630 MB on Replica, and about 12 frames per second against under 1 and 206 MB against 1,120 MB on ScanNet. Feature-3DGS still scores slightly higher on segmentation accuracy on one of them. The authors also state that the method depends on a good initial sparse reconstruction and can struggle outdoors, where that step tends to fail.
The paper was first posted in June and presented at ECCV 2026, and the code followed on September 28 as release v0.3.0. The release covers training and evaluation for both indoor datasets and ships the pretrained room model, which the team rechecked for image quality on an RTX 5090. The release notes include no speed measurements and do not reproduce the paper's full results tables, and the code requires Linux and a recent NVIDIA GPU.
The code is available on GitHub under the MIT license.
