Leveraging NeRF-Rendered Images for 3D Gaussian Splatting

A new method combines two competing image-synthesis techniques to render street scenes with both high speed and high quality, according to a paper submitted to arXiv on 8 June 2026 [1][2]. The approach uses images generated by a neural radiance field to train a 3D Gaussian splatting model, inheriting the strengths of each. Neural radiance fields (NeRF) and 3D Gaussian splatting (3DGS) are two mainstream approaches for novel view synthesis, the task of creating images from new camera angles [1][2]. The techniques often show complementary performance: 3DGS demonstrates faster rendering speed, while NeRF demonstrates higher rendering quality [1][2]. 3D Gaussian splatting, a volume rendering technique originally introduced in the early 1990s, was revitalized in 2023 when a research group from Inria proposed a version offering real-time radiance field rendering [3]. Multiple works soon followed, including extensions for dynamic scenes [3]. The proposed method targets street scenes and utilizes a pre-trained street-specific NeRF method to produce training images for a target 3DGS method [1][2]. During 3DGS training, the NeRF-rendered images serve two purposes: removing transient objects from street-level input views and generating bird's-eye views as additional training perspectives [1][2]. The process incorporates a diffusion-based image enhancement step to improve the quality of the additional views [1][2]. Experimental results on one synthetic and two real datasets demonstrate that the method improves street-scene rendering while preserving the speed of 3DGS and the quality of NeRF [1][2]. The paper was submitted through arXivLabs, a framework that allows collaborators to develop and share new features on the arXiv website [1].