Six Web 3DGS Renderers Benchmarked, With Procedure and Code Released Under MIT

What happened

According to the README, each model × backend combination was measured independently three times, for a total of 81 runs and 432 performance groups. The metrics are FPS and P50/P95 frame interval; memory and sort latency are defined separately. The site ranks rendering with depth sorting and stochastic rendering, which is close to unsorted, separately. A summary JSON and table CSV can also be downloaded from the site.

How it works

In the lab for testing your own data (/test/), you drop a PLY or SPZ file or specify a CORS-enabled URL. Files are processed on the device, and SPZ files are converted to PLY locally. You can choose the resolution (720p to 4K), the number of rounds, the measurement duration in seconds, and the camera, and export results as JSON or PDF.

By the numbers

Why it matters

The benchmark states explicitly that sort-key precision and the lag of asynchronous sorting can affect visual appearance. It notes that when nearby splats share the same key, the alpha-blending result can change. Image quality needs to be checked in addition to FPS.

How to try it

1. Following the GitHub README, run npm ci and npm run dev, then open /test/ locally
2. Load your own PLY or SPZ file and decide the resolution, number of rounds and sampling time
3. Run both the static and orbit tests, check FPS and P50/P95, and save the results as JSON/PDF
4. Compare using the same model, camera, physical resolution and browser settings. The README also cautions to keep these consistent

Caveats and open questions

• The results are values from a single Mac (Apple M5 Pro, Chrome 156). Rankings on Windows/NVIDIA, mobile and other browsers cannot be confirmed from the sources.
• The WebGPU sorting for PlayCanvas and SuperSplat Viewer is described as "4-bit radix on this Apple GPU", and may differ depending on the GPU.
• The page states explicitly that pipelines and the number of effective splats differ by implementation, so this is not an algorithm comparison on identical input.
• The MAE/PSNR in the image comparison is the difference from each backend's own sorted image. It is stated not to be a metric of image quality.
• Aholo Viewer is WebGL2 only (because its npm package has no public WebGPU renderer). Stochastic rendering applies only to the two backends PlayCanvas WebGPU and Gaussian Splat Lite.
• There are inconsistencies in the benchmark's own descriptions. The README mixes "Eight backends" and "nine backends", and Aholo Viewer is missing from the comparison list at the top. The GitHub release v1.0.0 description says "24 runs, 180 groups", which differs from the README's "81 runs, 432 groups". Which is the latest cannot be confirmed from the sources.
• At the time of publication, the GitHub repository is small, with 4 stars, and whether third parties have reproduced or verified it cannot be confirmed from the sources.
• Individual FPS values and rankings are not included in the source material, so this article does not state which renderer is better.

Glossary

3D Gaussian Splatting (3DGS)

A 3D representation that expresses a scene with many elliptical Gaussian distributions (splats) and renders it quickly.

Splat sorting

Processing that orders splats from far to near, or near to far, relative to the viewpoint before drawing, for transparency compositing.

Radix sort

An integer sort that orders keys digit (bit) by digit. Used on both GPUs and in Workers.

Stochastic rendering

A method that handles transparency through stochastic sampling instead of sorting. It produces noise, so it is often compensated for with post-processing.

WebGPU / WebGL2

APIs for using the GPU from a browser. WebGPU is newer and supports compute shaders.

SH3

Third-order spherical harmonics, coefficients that represent color changes depending on viewing angle. They increase the data size.

P50/P95

The median frame interval and the boundary of the slowest 5%. The latter is a guide to stutter.

Sources

Primary sources (publisher, paper, repository)

• 3DGS Renderer Bench · Retina3dgs-renderer-bench.vercel.app（3dgs-renderer-bench.vercel.app）・en
• 3DGS Renderer Bench（ブラウザラボ）3dgs-renderer-bench.vercel.app（3dgs-renderer-bench.vercel.app）・en
• WilliamLiu-1997/3dgs-renderer-benchGitHub（github.com）・en

The X post that surfaced this story

• View @lyzzz97's post on Xx.com

This article was generated in Japanese by AI from the sources listed (primarily primary sources) and published automatically, without prior human review, after automated checks such as source matching (how we use AI). Figures and specifications are as stated by the publishers and have not been independently verified by us. Please consult the original sources. To report an error, use our contact form; we will review it promptly and correct it. This page is an AI translation of the Japanese original. If anything differs, the Japanese version prevails. Japanese original