NeRF vs Gaussian Splatting: Which Radiance Field Method to Use
NeRF and gaussian splatting both turn photographs into a viewable radiance field, but they store it so differently that the choice decides training time, hardware, and whether anyone can edit the result. A nerf vs gaussian splatting comparison for real capture pipelines, not a benchmark chart.
A folder of photos or a two-minute walkthrough video is not yet a 3D scene. Somewhere between the capture and the link a client eventually opens, a radiance field method has to turn those pixels into something a browser can render from any angle, and for the last three years that choice has come down to nerf vs gaussian splatting. One is the neural approach that started this field in 2020. The other is the point-based method that displaced it in most production pipelines within a year of its own publication. Both solve the same problem, rebuilding a continuous view of a space from a set of photographs, but they store that reconstruction so differently that the choice shapes everything downstream: how long processing takes, what hardware plays the result back, and whether anyone besides the person who trained it can open the scene and actually change something in it.
NeRF vs gaussian splatting: the short answer
For nearly every capture-to-delivery pipeline built around a real space, gaussian splatting is the better default. It trains in minutes rather than hours, renders in real time without a GPU-hungry network running at display time, and produces a scene an editor can select, move or delete piece by piece. NeRF and its faster descendants such as Instant-NGP still earn a place where the deliverable is a continuous volumetric field for further computer vision work rather than a scene built to be walked, and Nerfstudio remains the more mature tool for exactly that job.
How we are comparing
The nerf vs gaussian splatting question is really six smaller trade-offs bundled together, and naming them before making any judgment is what keeps a comparison honest rather than a verdict dressed up as analysis.
- Training time. How long from a folder of captured images to a field you can actually view.
- Rendering method and hardware. What plays the result back: a network queried per pixel, or geometry a GPU can rasterize directly.
- Editability. Whether a person can open the finished scene and move a chair, delete a floater, or add a hotspot without retraining anything.
- Output portability. Whether the result travels as a small, self-contained file or stays tied to the framework that trained it.
- Behaviour on real capture. How each method holds up on a handheld walkthrough with imperfect overlap, not just a curated lab dataset.
- Tooling maturity. Whether there is an actively maintained pipeline behind the method, or a research repo someone has to keep patched.
NeRF, in short
A NeRF is a small neural network trained per scene to answer one question for any 3D point and viewing direction: what colour and density does the space have there. Rendering a pixel means marching a ray through the volume, querying that network at points along the ray, and integrating the results, the differentiable volume rendering scheme Mildenhall and colleagues introduced in their original 2020 paper. That original method optimized one network from scratch for each scene, and training was measured in hours even on a modest set of viewpoints.
Instant-NGP changed the arithmetic. Müller, Evans, Schied and Keller's 2022 paper replaced part of that slow per-scene optimization with a multiresolution hash encoding, and NVIDIA's own project page for the work reports training a usable NeRF on a single GPU in a matter of seconds, down from the hours the original approach needed. That is the version most people mean today when they say NeRF is fast, and it is the version behind most NeRF demos built since 2022.
Nerfstudio sits a layer above both. Tancik and a large group of Berkeley co-authors published it at SIGGRAPH 2023 as a modular PyTorch framework that wraps NeRF variants, Instant-NGP-style hash encodings included, behind one command-line interface, a real-time training viewer, and export paths to video, point cloud and mesh. It is the closest thing NeRF has to a production pipeline, and a team still running a NeRF workflow today is more likely running it through Nerfstudio than through a standalone research repo.
The trade for that continuous representation shows up the moment you try to display it. Producing a frame still means marching rays and querying the network at every sample point along every one of them, so even a NeRF that trained in seconds needs either a capable GPU sitting nearby or a separate baking step before it reaches interactive frame rates in a browser tab or on a phone.
Gaussian Splatting, in short
3D Gaussian Splatting drops the neural network at render time entirely. Kerbl, Kopanas, Leimkühler and Drettakis published the method in 2023, representing a scene as a few hundred thousand to several million 3D Gaussians, each one a small anisotropic blob carrying a position, a colour and an opacity. The Gaussians start from the sparse point cloud a structure-from-motion step like COLMAP already produces from the same photos, then get refined by gradient descent against the training images. Rendering is rasterization: project the Gaussians onto the frame and composite them, the same class of GPU operation that has drawn triangles for thirty years, aimed here at soft blobs instead.
The paper's own benchmark is the number that mattered to the field: real-time display at 1080p above 100 frames per second, with training landing in minutes rather than hours for scenes that reached quality comparable to the neural baselines it was measured against.
Being explicit instead of implicit changes two things downstream. Because a splat scene is a list of primitives rather than a black-box function, an editor can select one Gaussian, move it, delete it or recolour it without retraining anything, which is most of why tools like Captures Studio's Splat Editor and Multi Splat work at all: both operate directly on the Gaussians instead of on a network's weights. And because the format is just points with attributes, it travels. Files in .spz, .ply, .splat and .ksplat open in a browser, on a phone, or in a VR headset with no dedicated NeRF renderer installed anywhere.
The table below turns the nerf vs gaussian splatting comparison into concrete figures instead of adjectives.
| Criterion | NeRF (Instant-NGP / Nerfstudio) | Gaussian Splatting |
|---|---|---|
| Representation | Implicit: an MLP queried per sample point | Explicit: a list of anisotropic 3D Gaussians |
| Typical training time | Seconds to minutes with hash encoding; hours for the original 2020 method | Minutes, per Kerbl et al.'s own benchmark |
| Rendering | Ray marching with a network query per sample | Rasterization; 1080p above 100 fps per Kerbl et al., 2023 |
| Playback hardware | A capable local GPU, or a pre-baked export | Browser, phone, or VR headset, no install |
| Editability | Retrain or fine-tune to change the scene | Select, move, delete or recolour Gaussians directly |
| Output portability | Tied to the training framework unless exported | Open point-based files: .spz, .ply, .splat, .ksplat |
| Tooling maturity | Nerfstudio, maintained since 2023 | Cloud pipelines including Captures Studio, plus growing open viewers |
Where NeRF wins
NeRF's continuous, implicit field is still the right tool when the deliverable is not a scene to walk but a function to query. Extracting depth or surface normals from an arbitrary viewpoint, generating novel training views for a downstream vision model, or comparing against the sparse-view and dynamic-scene benchmarks that novel-view synthesis research has built around NeRF since 2020 all lean on a representation that returns a smooth answer everywhere in space, not just where a splat happened to land.
Nerfstudio's export pipeline also gives NeRF a foothold once the destination is a triangle mesh. A team already working downstream of a conventional 3D pipeline, feeding a game engine or a CAD tool that expects polygons, can go NeRF-in, mesh-out through Nerfstudio's exporters without introducing a splat format that pipeline was never built to read.
And it's worth saying plainly: Nerfstudio can now train gaussian splats too. Its Splatfacto method, built on the open-source gsplat rasterizer, is Nerfstudio's own implementation of the Kerbl et al. approach, initialized from the same COLMAP points a NeRF run would use. For a team already fluent in Nerfstudio's CLI and viewer, that is a real reason to stay rather than start over on a separate stack, even after choosing splatting as the output representation.
Where Gaussian Splatting wins
Everything that makes gaussian splatting the practical default shows up once a scene has to leave a research machine and reach an actual viewer. Real-time rendering with no network in the render loop means the same scene plays on a laptop, a phone, or a VR headset without a baking pass in between. Direct editability means a hotspot, a camera path, or a deleted floater is a few clicks, not a retraining run. Open point-based formats mean the file itself, not a bundle of framework code, is the deliverable.
Captures Studio builds its whole pipeline around that difference. Processing happens in the cloud in 10 to 120 minutes with no local GPU and nothing to install, and it accepts whatever a team already has: ordinary phone or DSLR photos, plain video in MP4 or MOV, 360 footage from an Insta360, Ricoh Theta or GoPro Max, or a LiDAR point cloud in E57, LAS, LAZ or PLY. The result opens as .spz, .ply, .splat or .ksplat and ships as a link, an embed, or a QR code, on a phone, a tablet, a desktop or a headset, with no app to install on the other end. A point cloud from a scanner stays an accurate but sparse diagram; running the same captured data through image-to-3d or lidar-to-3d reconstruction turns it into continuous, photoreal surfaces a client walks instead of interprets.
Gaussian splatting also degrades more gracefully on the capture a real project actually produces: a walkthrough with a few rushed corners and inconsistent lighting, rather than the evenly lit multi-view rig a benchmark dataset assumes. Since the Gaussians are optimized directly against whatever photos exist, a weak region shows up as a soft or noisy patch that an editor can clean up by hand, not as a network that quietly failed to converge somewhere off camera.
Which to choose
- The deliverable is a shareable, walkable scene a client opens on their own device. Choose gaussian splatting, and choose a cloud pipeline over a self-hosted trainer if speed to delivery matters more than owning the GPU.
- The work is computer vision research needing a continuous, differentiable field: depth supervision, view synthesis benchmarking, or anything measured against NeRF-era datasets. Keep NeRF, and run it through Nerfstudio rather than the 2020 reference code.
- The source material mixes photos, video, 360 footage and LiDAR across one project. A pipeline built for varied capture types, rather than a separate NeRF configuration per input, avoids retraining your process for every new format.
- A team is already fluent in Nerfstudio and only needs internal previews, not client-facing scenes. There's no urgent reason to migrate that internal workflow. Splatfacto even lets that same team try gaussian splatting inside the tool they already know before deciding whether a dedicated pipeline is worth adopting.
- The scene needs to be triggered from an app or an internal tool rather than opened by hand. A splatting pipeline exposed through the reconstruction API fits into an existing upload flow far more easily than a Nerfstudio job queued on a research machine.
FAQ
Is gaussian splatting a type of NeRF?
No, though the confusion is common. "NeRF" specifically names the implicit, neural-network representation Mildenhall and colleagues published in 2020. Gaussian splatting, introduced by Kerbl and colleagues in 2023, is explicit and point-based, with no neural network involved at render time. Both fall under the broader term "radiance field method," which is why the two get compared so often, alongside the older photogrammetry pipeline both were built to replace (see our gaussian splatting vs photogrammetry breakdown for that comparison).
Does Nerfstudio train gaussian splats too?
Yes. Its Splatfacto method trains gaussian splats using the open-source gsplat rasterizer, initialized from the same COLMAP structure-from-motion points a NeRF run in Nerfstudio would use. A cloud pipeline built specifically around splatting from the start still tends to be faster from raw capture to a finished, shareable scene, since it skips setting up the framework locally.
Do I still need COLMAP for gaussian splatting?
Most implementations, the original Kerbl et al. method included, start from sparse points a structure-from-motion tool like COLMAP produces, so a capture still needs enough image overlap for pose estimation to succeed before splatting can begin. Cloud services vary in whether that step is exposed directly or runs invisibly as part of processing.
A scan that took forty minutes to walk is, in the end, just a link someone opens on a phone. Which radiance field method got it there matters far less to the person opening that link than whether the result loads instantly, holds up when they turn a corner, and lets someone fix the one part that came out wrong. That is the test both methods are actually being measured against now, whatever a benchmark table says.
All trademarks belong to their respective owners. No affiliation, partnership or endorsement is implied.