Neural Scene Flow Fields

Neural Scene Flow Fields (NSFF) represents a seminal architecture in the evolution of Neural Radiance Fields (NeRF) specifically designed for dynamic, non-rigid scenes. Unlike static NeRF, which fails in the presence of motion, NSFF represents a scene as a continuous function of both space and time (4D). The architecture utilizes a coordinate-based Multi-Layer Perceptron (MLP) to map 3D position and temporal instances to volume density and RGB color, but critically introduces a scene flow field. This field predicts 3D motion vectors (forward and backward flow) between adjacent frames, enabling the model to establish temporal correspondences. By integrating monocular depth estimation priors and consistency losses, NSFF can reconstruct complex human movements and fluid dynamics from a single moving camera. As of 2026, it remains a foundational framework for researchers and VFX engineers building temporal view synthesis tools. The model's ability to decouple static and dynamic elements allows for sophisticated video editing, such as time-freezing (bullet time) or moving-camera slow motion, positioning it as a core technology for the next generation of generative video and spatial computing pipelines.