Neural Sparse Voxel Fields (NSVF) is a novel neural scene representation designed for fast and high-quality free-viewpoint rendering. Unlike traditional methods that rely on 3D geometry and appearance models, NSVF uses a set of voxel-bounded implicit fields organized in a sparse voxel octree to model local properties in each cell. This approach allows for efficient rendering by skipping voxels containing no relevant scene content, significantly reducing computation time. NSVF can be learned from posed RGB images using a differentiable ray-marching operation, and it outperforms state-of-the-art methods like NeRF in terms of both speed and quality. The method is also applicable to scene editing and composition, and has been evaluated on various challenging tasks, including multi-scene learning, free-viewpoint rendering of dynamic and indoor scenes, and large-scale scene rendering. NSVF provides a more efficient and detailed representation of 3D scenes, enabling higher visual quality and faster rendering compared to existing methods.Neural Sparse Voxel Fields (NSVF) is a novel neural scene representation designed for fast and high-quality free-viewpoint rendering. Unlike traditional methods that rely on 3D geometry and appearance models, NSVF uses a set of voxel-bounded implicit fields organized in a sparse voxel octree to model local properties in each cell. This approach allows for efficient rendering by skipping voxels containing no relevant scene content, significantly reducing computation time. NSVF can be learned from posed RGB images using a differentiable ray-marching operation, and it outperforms state-of-the-art methods like NeRF in terms of both speed and quality. The method is also applicable to scene editing and composition, and has been evaluated on various challenging tasks, including multi-scene learning, free-viewpoint rendering of dynamic and indoor scenes, and large-scale scene rendering. NSVF provides a more efficient and detailed representation of 3D scenes, enabling higher visual quality and faster rendering compared to existing methods.