An overview of NeRF and Gaussian Splatting

Computer vision has come a long way since its inception in the mid-1960s, and neural radiance fields (NeRFs) are a large part of the contemporary story.

 

Neural radiance fields (NeRFs) are a powerful form of artificial intelligence (AI) that utilizes 2D images to synthesize a 3D scene and “learn” what a realistic 3D representation of the base information would look like. While NeRFs initially revolutionized computer vision in the early 2020s, a new method called Gaussian Splatting quickly overtook NeRFs. In this comprehensive review, researchers explore some of the differences between these two methods and their applications in the hopes of better informing future computer vision practitioners.

 

Researchers put forth their review in Computational Visual Media in May 2026.

 

The field of computer vision has been drastically changed when NeRFs came into play; their applications in urban mapping, autonomous navigation, virtual reality and augmented reality, to name a few, allowed for a more intelligent system and realistic representation to take place. While NeRFs had this corner of the tech field occupied, shortly after its emergence came a competitor, Gaussian splatting.

 

Gaussian splatting emerged in 2023 and quickly overtook NeRF as the primary framework, thanks to the faster training and inference time while maintaining high-quality imaging. Both of these technologies use novel view synthesis methods, a process where a limited number of 2D photos are taken from set camera angles, which are then processed through these frameworks to generate new, photorealistic images from angles not included in the original input. The difference lies in the representation and rendering paradigm used.

 

Implicit neural rendering methods use “hidden” 3D representations, like a pixel, and replace it with a continuous mathematical function to represent the image or scene. This implicit method is how NeRFs operate, and though the image quality might be lacking compared to the Gaussian technique, it is how file sizes are able to stay compact and modeling continuous. Explicit neural renderings store their representations in data structures, which is how the Gaussian splatting method goes. Explicit neural renderings generate larger files, but the image quality is unmatched.

 

“Although the research interest of the computer vision community has shifted towards Gaussian splatting in many key research areas, there remains much interest in NeRF and implicit/hybrid neural field rendering in applications where such a representation, or a volumetric rendering approach, is advantageous,” said Jonathan Li, author and researcher.

 

While Gaussian splatting is preferred for works where high image quality and real-time rendering matter, NeRFs remain an excellent choice for memory and storage efficiency. Hybrid methods aim to achieve the best of both worlds by combining the explicit rendering nature of Gaussian splatting along with the neural field representations, view-dependency, and expressivity of NeRFs. For applications such as 3D human avatars, a hybrid method is the more advantageous route to take, thanks to the quick rendering of Gaussian splatting and the continuous modeling of NeRFs.

 

There is space for both methods in the field of computer vision, especially when hybrid models are considered. Since the initial paper on NeRFs came out, work has focused on improving speed and image quality, as well as addressing some of the technique's other initial weaknesses.  Researchers hope this work will inform current and future computer vision practitioners about the strengths of each method and the scenarios where one method might be advantageous over another.

 

Kyle Gao, Dening Lu, Hongje He and Jonathan Li of the Department of Systems Design Engineering at the University of Waterloo, Yina Gao of the Faculty of Engineering at the University of Toronto and Linlin Xu of the Department of Geomatics Engineering at the University of Calgary contributed to this research.

 

The Natural Sciences and Engineering Research Council of Canada Discovery Grant supported this research.

 

DOI Link:

https://doi.org/10.26599/CVM.2026.9450535

 

---

---