Authors - Janitha Prabodha Bandara Dissanayaka, Pumudu A. Fernando, Manul Randula Singhe Abstract - Three-dimensional scene representation has moved quickly from Neural Radiance Fields (NeRF) to explicit volumetric methods such as 3D Gaussian Splatting (3DGS). 3DGS can render photorealistic scenes in real time, but editing these scenes is still difficult because Gaussian primitives do not have a fixed mesh structure or explicit topology. This becomes more challenging in dynamic scenes, where edits must remain stable across time and viewpoint. This paper presents a systematic literature review of neural volumetric editing methods for dynamic 3DGS and related NeRF-based representations. The review follows a PRISMA-guided process and analyses studies published between 2023 and 2026. The selected methods are grouped into three main paradigms: text-guided editing, interaction-based editing, and physics-based simulation. The review compares these methods using control precision, rendering speed, memory usage, editing time, temporal stability, and practical limitations. The findings show that text-guided methods are easy to use but often suffer from weak localisation and temporal inconsistency. Interaction-based methods provide stronger geometric control but struggle with complex topology changes. Physics-based methods produce more realistic motion but require higher computation and reliable material parameters. The review identifies the consistency gap, including flickering and texture swimming, as the main barrier to real-world dynamic neural volumetric editing.
Open Zoom