William Liu

Building 3D Web and WebGIS tools with Three.js, CesiumJS, 3D Tiles, and Gaussian Splatting.

[William Liu @lyzzz97]

Gaussian-Splat-Lite (#ThreeJS 3DGS Renderer) now supports SOG v3: opacity > 1, SH4, and Blind data

This includes opacity values above 1, degree-4 spherical harmonics (SH4), and Blind data loading, querying, and editing.

https://t.co/N2RDvilpfZ
https://t.co/DinsIW99P1

[William Liu @lyzzz97]

@andybak https://t.co/P6lXyOSO3T
Here you can do a simple test yourself

[William Liu @lyzzz97]

Benchmarking major web 3DGS renderers: Spark, Radix, PlayCanvas, SuperSplat, GSL and Aholo

Setup: Apple M5 Pro, 48 GB memory, Chrome 156, 3024 × 1964 pixels.

Detailed Report: https://t.co/jO9Xv38crV
Test your own model: https://t.co/P6lXyOSO3T
GitHub: https://t.co/f4Hd1IA7gM

[William Liu @lyzzz97]

@SumeruAI Yeah, RAD is the most complicated, but it has the best visual quality

[William Liu @lyzzz97]

Just added LCC2 streaming to Gaussian-Splat-Lite 🚀

Smooth LOD crossfades as the camera moves.

GSL now supports most of LOD streaming formats, including RAD, SOG, LCC2, and 3D Tiles.

https://t.co/N2RDvilpfZ

#GaussianSplatting #Threejs #3DGS

[William Liu @lyzzz97]

Three.js: Gaussian Splat Lite · Spark · Radix
PlayCanvas: PlayCanvas Engine · SuperSplat Viewer
Other: Aholo Viewer

[William Liu @lyzzz97]

@zyonbuilds Yes. When streaming from a server, downloading typically accounts for over 85% of the loading time. So a 55% reduction in file size can translate to over 40% faster loading in the browser.

[William Liu @lyzzz97]

SH codebooks are now supported in the Node.js 3DGS → RAD converter in Gaussian-Splat-Lite, my #Threejs/WebGPU 3DGS renderer!

On a 60M-splat scene:
• 55.6% smaller files
• Full conversion, including SH clustering: 11m vs Spark’s 11h 20m (~58× faster)

https://t.co/sVPJDxTqnk

[William Liu @lyzzz97]

That would be great, thanks! Having it integrated into splat-transform would be really useful.

This approach may not match NanoGS in PSNR, I’ve found that it preserves more of the important texture and structural details in many of the industrial building scenes I work with.

In my own LOD builder, this approach took about 30 minutes on an M5 Pro MacBook Pro to generate a complete multi-level LOD hierarchy from a 45 GB PLY containing 180M SH3 splats, with an 8 GB memory budget.

I think there’s still room for optimization. Once it’s integrated, I’d be happy to explore ways to improve memory usage and processing performance.

[William Liu @lyzzz97]

@slimbuck7 @playcanvas Yeah just @ you in discord

[William Liu @lyzzz97]

@slimbuck7 @playcanvas https://t.co/i3Ci2lUrh6
This is the result, the JastrebiaGora and OsadaSlavotovo corresponds to https://t.co/wuJVPmxlBA and https://t.co/EhPSSlpqEc you can compare

[William Liu @lyzzz97]

@slimbuck7 @playcanvas About 2 months ago

[William Liu @lyzzz97]

@slimbuck7 @playcanvas Already! In the discord!

[William Liu @lyzzz97]

Trained a 1M Gaussian splat model from Blender’s “The Junk Shop” scene. SH3 and SH0 versions below 👇

SH3 https://t.co/RCrLIZyPT0

SH0 https://t.co/hQy6Xn93eS

[William Liu @lyzzz97]

@terralunatic These are open data from https://t.co/6J037cWWbp

[William Liu @lyzzz97]

Node.js tool to convert large 3DGS scenes to Spark’s RAD streaming format is now supported in Gaussian-Splat-Lite (GSL), my #Threejs/WebGPU 3DGS renderer

259M splats: ~25 min, 3.94 GiB peak RSS.
2.68–3.79× faster than Spark’s builder in 3 tests.

https://t.co/sVPJDxTqnk

[William Liu @lyzzz97]

@The1Broom It uses disk-backed paging for LOD construction and writes RAD chunks incrementally, so the full source and output aren’t held in RAM together. The 3.94 GiB is peak process RSS, including workers and buffers. The memory budget is configurable.