Open-vocabulary 3D Gaussian segmentation typically binds semantic features to each Gaussian primitive through per-scene gradient training, which can take one to four hours, or via more recent closed-form distillation. Both approaches commit each primitive to a single semantic label, which may limit flexibility in complex scenes.

The new method, called DIGS (Distribution-Informed Gaussian Splatting), instead leverages distribution information to perform segmentation without any training. This makes it a training-free alternative that avoids the time cost and the rigid one-primitive-one-label constraint of prior work.

The abstract is brief, so details on the exact mechanism and evaluation are not yet available. However, the core contribution is clear: a distribution-informed, training-free approach to open-vocabulary 3D segmentation that departs from the training-heavy paradigm of existing methods.