So i vibe coded a threejs webgpu splat renderer from scratch in an evening with gpt 5.4 Introducing Spark++ Give it a week or two, performance is slightly worse on fast camera movements but 60 fps on my toaster in less dense areas. Had to reduce overall density for... screen recording, but it works on webgpu! And I have absolutely no clue how Splats work, zero Demo: Repo:show more

Amazing. Keep it up! Make sure it support LoD and streaming. Target size, 100 million splats. That's when things get difficult.

it has LOD, no streaming, what is target size? It wouldn't ever render 100 million, it slices the entire splat into lots of small pages and only renders what's visible in the camera

Yeah, so LoD generally builds an LoD tree where splats close by are high resolution, and those far away are lower resolution. Generally this is done by having the highest resolution be leafs on a tree, which aggregate up and create blended colors (and increased size) as you go. You can set a splat budget (e.g. 2 million) for being on the screen at all times. However, they're all stored in memory since you're simply saving GPU time. Streaming is where you normally deal with memory and bandwidth contention by carving the scene into pages, and then loading them incrementally. This saves you loading time (e.g. you can see the scene immediately) and you can fix total memory usage by evicting pages that are too far away. Here is an example of a 16million splat scene with LoD and streaming. It also has a proxy mesh that takes a moment to load:

isn't building, sorting and filtering the tree taking up massive cpu time?

Building the LoD tree(s) is generally done offline. But yeah, for traversal, sorting, multi-views etc it's really hard to get it performant. In particular if you have multi-device support. For example Spark supports VR and mobile. Spark is currently at or ahead of the state of the art. So there really isn't any literature or prior code for the models to draw from. It gets even harder if you want to implement portals. Because now you need multiple multiple LoD trees and clipping which has to happen with shaders to be performant. And all that has to interact with the streaming / paging system.

do u have a large splat for me

Yeah. Will send in a bit

🗿🖤

Nice

banger

just vibin lately

the only way

You got further than me using claude :) Claude just messed up all compute shaders of the base project so left. Will try codex on my claude wip and see if I can get it to work.

cracked 🔥

Bro what, how the..

Nice, I also vibe coded one:

Rad work.

lovely

I’ve coded splats viewer that renders in symbols in a same way 👀

What the hell I worked with SparkJS quite a bit and always considered it basically magic, especially the 2.0 version which has LODs (so you can limit the amount of splats drawn to n)

it is magic still

Right now you just have a point cloud (a dark one at that) so they are not actual splats yet. You need to get the ellipsoid shape, scale and rotation working, which prob won’t effect render speed, just decoding.

check the link i've updated it since

AI trap... you did not know how it work, now you have a thing, but still does not know how it works..

60fps on a toaster with webgpu splats? That's actually wild. The real question: did GPT 5.4 figure out the math or just copy-paste from threejs examples? Either way, the fact that it works at all is the real flex here.