This paper introduces PlenOctrees, an octree-based 3D representation that enables real-time rendering of Neural Radiance Fields (NeRFs). The method allows rendering of 800×800 images at over 150 FPS, which is more than 3000 times faster than conventional NeRFs. The approach preserves the ability of NeRFs to perform free-viewpoint rendering of scenes with arbitrary geometry and view-dependent effects. Real-time performance is achieved by pre-tabulating the NeRF into a PlenOctree. To preserve view-dependent effects such as specularities, the appearance is factorized via closed-form spherical basis functions. The method trains NeRFs to predict a spherical harmonic representation of radiance, removing the viewing direction as an input to the neural network. Furthermore, PlenOctrees can be directly optimized to minimize the reconstruction loss, leading to equal or better quality compared to competing methods. The octree optimization step also reduces training time. The real-time neural rendering approach enables new applications such as 6-DOF industrial and product visualizations, as well as next-generation AR/VR systems. PlenOctrees are amenable to in-browser rendering. The project page provides an interactive online demo, as well as video and code: https://alexyu.net/plenoctrees. The method is evaluated on standard benchmarks with scenes and objects captured from 360-degree views, demonstrating state-of-the-art performance in image quality and rendering speed. The interactive viewer enables operations such as object insertion, visualizing radiance distributions, decomposing the SH components, and slicing the scene. The contributions include the first method achieving real-time rendering of NeRFs with similar or improved quality, NeRF-SH, a modified NeRF trained to output appearance in terms of spherical basis functions, PlenOctree, a data structure enabling efficient view-dependent rendering of complex scenes, and an accelerated NeRF training method using early training termination followed by direct fine-tuning on PlenOctree values. The method is compared with other approaches, including neural sparse voxel fields (NSVF) and AutoInt, and shows significant improvements in speed while maintaining image quality. The method is also shown to accelerate NeRF training by allowing early termination and fine-tuning on PlenOctree values. The results demonstrate that the approach can accelerate NeRF-based rendering by 5 orders of magnitude without loss in image quality. The method is implemented for real-time and in-browser applications, enabling interactive viewing of PlenOctrees in the browser. The method is also shown to enable virtual online stores in VR, where any products with arbitrary complexity and materials can be visualized in real-time while enabling 6-DOF viewing. The method has limitations, including a larger memory footprint compared to the original NeRF model and the need for further work to apply it to unbounded and forward-facing scenes. The method has potential to become aThis paper introduces PlenOctrees, an octree-based 3D representation that enables real-time rendering of Neural Radiance Fields (NeRFs). The method allows rendering of 800×800 images at over 150 FPS, which is more than 3000 times faster than conventional NeRFs. The approach preserves the ability of NeRFs to perform free-viewpoint rendering of scenes with arbitrary geometry and view-dependent effects. Real-time performance is achieved by pre-tabulating the NeRF into a PlenOctree. To preserve view-dependent effects such as specularities, the appearance is factorized via closed-form spherical basis functions. The method trains NeRFs to predict a spherical harmonic representation of radiance, removing the viewing direction as an input to the neural network. Furthermore, PlenOctrees can be directly optimized to minimize the reconstruction loss, leading to equal or better quality compared to competing methods. The octree optimization step also reduces training time. The real-time neural rendering approach enables new applications such as 6-DOF industrial and product visualizations, as well as next-generation AR/VR systems. PlenOctrees are amenable to in-browser rendering. The project page provides an interactive online demo, as well as video and code: https://alexyu.net/plenoctrees. The method is evaluated on standard benchmarks with scenes and objects captured from 360-degree views, demonstrating state-of-the-art performance in image quality and rendering speed. The interactive viewer enables operations such as object insertion, visualizing radiance distributions, decomposing the SH components, and slicing the scene. The contributions include the first method achieving real-time rendering of NeRFs with similar or improved quality, NeRF-SH, a modified NeRF trained to output appearance in terms of spherical basis functions, PlenOctree, a data structure enabling efficient view-dependent rendering of complex scenes, and an accelerated NeRF training method using early training termination followed by direct fine-tuning on PlenOctree values. The method is compared with other approaches, including neural sparse voxel fields (NSVF) and AutoInt, and shows significant improvements in speed while maintaining image quality. The method is also shown to accelerate NeRF training by allowing early termination and fine-tuning on PlenOctree values. The results demonstrate that the approach can accelerate NeRF-based rendering by 5 orders of magnitude without loss in image quality. The method is implemented for real-time and in-browser applications, enabling interactive viewing of PlenOctrees in the browser. The method is also shown to enable virtual online stores in VR, where any products with arbitrary complexity and materials can be visualized in real-time while enabling 6-DOF viewing. The method has limitations, including a larger memory footprint compared to the original NeRF model and the need for further work to apply it to unbounded and forward-facing scenes. The method has potential to become a