Donovan Hutchence

I am still fascinated by our whacky "Speed of light" Gaussian Splatting renderer.

It depth culls away nearly 90% of the scene and renders in 4ms. It's wild! and also completely unusable.

My home-brew 4dgs pipeline is slowly improving.

Individual Gaussians are still too wobbly (technical term). I might try post-process the frame data to filter/smooth out some of the noise.

This scene has 10k gaussians, 5k of them on the face. Gaussians are tracked between frames, which is very useful when it comes to compressing the data.

The source footage for this clip is my own rig of 10x Raspberry PI global shutter cameras (1.4MP each). I'm not using GenAI for in-between frames.

I really need new test footage, desperately... cc @willeastcott! :D

This illustrates why we're adding support for alpha > 1 to SOG v3.

Under repeated rounds of decimation, alpha compensation prevents gaps from forming.

For this test, the background is set to 🔴 BRIGHT RED 🔴 to highlight gaps between gaussians without compensation.

I'm working on the @playcanvas SOG v3 proposal.

So far we have:
- Tighter data packing
- Up to 4 bands of SH
- Alpha > 1 support (compensation post-decimation)
- Per-gaussian custom/blind data
- meta.json extras field, à la glTF

Would love your feedback!👇
github.com/playcanvas/spl…

I might be having a little too much fun working on our @Apple Vision Pro support...

Church on a stick!

Scene by @Schindelar3D at superspl.at/scene/c67edb74