Technical

Frankensplat Trains Gaussian Splats Inside a Browser Tab: What WebGPU Makes Possible and What the Benchmark Does Not Show

ยท RenderBob team

Rafael Spring's Frankensplat trains Gaussian splats from COLMAP reconstructions in a browser tab with WebGPU. The speed and quality figures are the developer's.

A browser-shaped glass frame holds a vase reconstruction, with camera pyramids and a point cloud on the left becoming colored Gaussian ellipsoids and a rendered still life on the right, above a GPU tile and beside an empty stopwatch.

Rafael Spring has released Frankensplat, a tool that trains Gaussian splats from COLMAP camera reconstructions inside a browser tab, using JavaScript and WebGPU. Digital Production covered it on 18 November. That article did not test the tool. The benchmark below is the developer's, including the fixed-resolution detail on the developer's own study page.

What it does

  • Input. Folders or ZIPs holding images and a COLMAP sparse reconstruction. The photogrammetry step is a separate tool by the same developer. A combined photos-to-splats workflow is not available yet, though the developer says the two projects may merge.
  • Output. PLY files, for viewing and export.
  • Controls. Splat budget, image scale, PPISP appearance correction and progressive resolution, which is on by default. Checkpoints resume against the matching original dataset.
  • Cost. Digital Production describes it as free to use and quotes no licence terms.

Requirements

Training needs WebGPU, sufficient GPU limits and enough GPU memory. Viewing uses WebGL2. The benchmark machine had an RTX 3070 with 8 GiB of VRAM, on Windows with Edge. No minimum specification is given. The site and its modules must load online first, and later offline use depends on browser caching.

What the benchmark says, and does not say

The developer compared Frankensplat with a development build of LichtFeld Studio, which uses CUDA, in a study of 26 single runs.

  • Across five matched scenes, fixed-resolution Frankensplat took 30.3 percent less time than LichtFeld with detail features disabled. Progressive mode is the default, and that 30.3 percent figure is not the progressive timing.
  • Across seven scenes, the progressive version averaged 28.52 dB fitted PSNR against 29.39 dB for LichtFeld with detail off. That is slightly lower quality.
  • The metric is appearance-fitted PSNR. Digital Production does not present it as an independent held-out score, and the speed result has not been independently reproduced.
  • Digital Production lists no other limits.

A speed gain came with a small quality gap in the developer's own test, and one test on one card does not predict results on yours.

Why a browser matters

Browser tools remove installation steps and can run on a machine where you cannot install software. They also depend on browser support and on the GPU limits a browser exposes, which are tighter than a native app's.

How to test a tool like this

  • Use a capture set you already know, with a splat you trust for comparison.
  • Note your GPU and VRAM, since the benchmark gives no minimum.
  • Check the PLY in the viewer you will actually use downstream.
  • Time a full run, including loading, not just training.
  • Repeat once. The benchmark used single runs.

What to watch

Independent tests on other GPUs and browsers, a combined photos-to-splats workflow, and published licence terms.

RenderBob 3 and this story

RenderBob 3 chooses local Apple Silicon or cloud generation per clip. A browser-local splat trainer puts the same choice in front of a 3D artist: run on your own machine, or somewhere else, depending on the job.

More from the blog

All posts