NeRF Viewer Online — View Neural Radiance Fields in Browser

What Are Neural Radiance Fields?

Neural Radiance Fields, universally known as NeRFs, represent one of the most significant breakthroughs in 3D computer vision since the invention of photogrammetry. Introduced in a landmark 2020 paper by Mildenhall et al., NeRFs learn to represent a 3D scene as a continuous volumetric function using a neural network. You feed the network a collection of photographs taken from different angles, and it learns to synthesize novel views of that scene from any arbitrary camera position.

The core idea is elegant: a neural network takes a 5D input (3D spatial coordinates plus 2D viewing direction) and outputs a color and density value for that point in space. By sampling points along camera rays and accumulating these color and density values through volume rendering, the network produces photorealistic images of the scene from viewpoints it has never seen. The results, especially for complex scenes with reflections, translucency, and fine geometric detail, can be strikingly lifelike.

Since the original paper, the field has exploded. Hundreds of variants have been published: Instant-NGP from NVIDIA dramatically sped up training from hours to seconds using hash encoding. Mip-NeRF improved quality by reasoning about the volume of each pixel rather than a single ray. Nerfacto, the default model in Nerfstudio, blends the best ideas from multiple papers into a practical all-rounder. Block-NeRF scaled the technique to entire city blocks. The pace of innovation has been relentless.

The Viewing Problem

Despite NeRFs producing stunning results, actually viewing them remains surprisingly difficult. A trained NeRF model is not a mesh, not a point cloud, and not a traditional 3D file you can open in Blender or upload to a website. It is a neural network checkpoint, a set of learned weights that must be run through a specific inference pipeline to produce images.

To view a NeRF interactively, you typically need one of the following tools:

• NVIDIA Instant-NGP ships with a real-time viewer that renders the NeRF directly on an NVIDIA GPU using custom CUDA kernels. It runs only on machines with compatible NVIDIA hardware and requires building from source on Linux.
• Nerfstudio provides a browser-based viewer that streams rendered frames from a GPU server. While it technically runs in the browser, it still requires a powerful GPU machine running the Nerfstudio backend. The browser is just a thin client.
• Luma AI offered a cloud-hosted NeRF viewer for a time, but their focus has shifted to Gaussian Splatting.

None of these solutions let you simply open a file in a browser and explore it locally. The fundamental issue is that volume rendering a neural network in real time requires significant GPU compute, something that today's WebGL and even WebGPU implementations cannot match for the original NeRF representation.

Mesh Extraction: Bridging NeRFs to the Web

The practical solution for bringing NeRF reconstructions to the web is mesh extraction. Both Instant-NGP and Nerfstudio support exporting the learned scene as a traditional polygon mesh, typically in OBJ or GLB format, using marching cubes or similar isosurface extraction algorithms.

The process works by evaluating the NeRF's density field on a regular 3D grid and extracting the surface where density crosses a threshold. Vertex colors or a baked texture atlas capture the appearance. The resulting mesh can be opened in any standard 3D viewer, including browser-based tools like GeometryViewer.

There are trade-offs. Mesh extraction loses some of the view-dependent effects that make NeRFs special, like reflections and translucency that change with viewing angle. The mesh resolution is limited by the grid resolution you choose, and higher resolutions produce enormous files. Texture baking introduces seams and resolution limits. But for the vast majority of use cases, an extracted mesh is perfectly adequate and infinitely more portable than the raw neural representation.

How to Export a Mesh from Nerfstudio

In Nerfstudio, the command is straightforward. After training your model, run the export command specifying your desired output format and resolution. You can choose OBJ with vertex colors, or GLB with a baked texture. A resolution of 512 to 1024 on the marching cubes grid is usually a good balance between quality and file size. The exported file can then be opened directly in GeometryViewer or any other 3D viewer.

How to Export from Instant-NGP

NVIDIA's Instant-NGP includes a mesh export option in its interactive viewer. You set a density threshold and grid resolution, then export to OBJ. The threshold controls how much of the reconstructed volume is included in the mesh. Lower thresholds capture more detail but also more noise. Some experimentation is usually needed to find the sweet spot for your particular scene.

Gaussian Splatting: The Faster Alternative

While NeRFs were still evolving, a newer technique arrived in 2023 that has largely overtaken them for real-time viewing: 3D Gaussian Splatting. Instead of representing a scene as a continuous neural field, Gaussian Splatting represents it as millions of tiny 3D Gaussian ellipsoids, each with a position, covariance (shape), color, and opacity.

The critical advantage of Gaussian Splatting for viewing is that it can be rasterized rather than ray-marched. Rasterization, projecting and sorting primitives onto a 2D screen, is exactly what GPUs are designed for. This means Gaussian Splats can be rendered at 100+ FPS on consumer hardware, and crucially, they can be rendered in WebGL browsers without a powerful backend server.

Several browser-based Gaussian Splatting viewers already exist. They load a .ply or .splat file containing the Gaussian primitives and render them in real time using WebGL2 or WebGPU. The visual quality rivals the original NeRF, complete with view-dependent effects, and the rendering is entirely client-side.

For many practitioners, Gaussian Splatting has replaced NeRFs entirely. The training is faster, the rendering is faster, and the results are comparable or better. However, NeRFs retain advantages for certain tasks like surface reconstruction, where a continuous density field produces cleaner meshes than converting splats to polygons.

How GeometryViewer Fits In

GeometryViewer handles the final, portable stage of the NeRF pipeline. Once you have exported your NeRF reconstruction as an OBJ or GLB mesh, you can drag and drop it into GeometryViewer to inspect it, share it, or embed it on a webpage. There is no software to install, no GPU server to maintain, and no compatibility issues to debug.

This is particularly useful for sharing NeRF results with collaborators or clients who do not have the technical setup to run Nerfstudio. Instead of asking someone to install Python, PyTorch, and CUDA, you export the mesh, share a link, and they view it on any device.

Current State of Browser-Based NeRF Rendering

True in-browser NeRF rendering, where the neural network runs client-side, remains experimental. A few research projects have demonstrated small NeRFs running in WebGL using baked representations like PlenOctrees or Sparse Neural Radiance Grids (SNeRG), which precompute and cache the neural network's output into a data structure that can be queried without running inference. These achieve real-time frame rates but are limited to small scenes and require a lengthy precomputation step.

WebGPU, the successor to WebGL, opens the door to running compute shaders in the browser, which could eventually enable more direct NeRF inference. But WebGPU adoption is still in progress, and the performance gap between browser-based compute and native CUDA is still large. For the foreseeable future, mesh extraction and Gaussian Splatting remain the practical paths to browser-based NeRF viewing.

Future Directions

The field is moving fast. Several trends are worth watching. First, hybrid representations that combine the editability of meshes with the photorealism of neural rendering are gaining traction. NeuMesh and similar projects attach neural features to mesh vertices, allowing real-time rendering of view-dependent effects on standard mesh infrastructure. Second, compression of Gaussian Splats is improving rapidly, with file sizes dropping from hundreds of megabytes to single digits without visible quality loss. Third, the line between NeRFs and Gaussian Splatting is blurring as researchers find ways to convert between representations and combine their strengths.

For web-based 3D viewing, the practical takeaway is clear: export your NeRF or Gaussian Splat scene as a mesh, and you can view it anywhere. The neural magic happens during reconstruction. The viewing can be simple, standard, and universal.