Optimization possibilities?

[botaohu]

Testing on Meta Quest

We used gsplat-unity to run the 500k ply file (downloaded from Marble) on Meta Quest compiled with Unity 6—achieving only 20 fps.

We can use the Meta Quest Browser to view https://marble.worldlabs.ai/.
It can view 500k low-res gsplat at 72 fps smoothly. Marble is implemented in spark.js.

Is it possible to optimize gsplat-unity to have spark.js-like high performance?

I think spark.js uses stochastic splat rendering.

https://github.com/sparkjsdev/spark/blob/main/src/shaders/splatVertex.glsl
https://github.com/sparkjsdev/spark/blob/main/src/shaders/splatFragment.glsl

https://arxiv.org/pdf/2503.24366

[wuyize25]

I don’t think Marble uses stochastic alpha blending. Its speed is probably because it offloads the sorting stage to the CPU, which reduces the workload on the GPU and allows the sorting to be amortized across multiple frames (if you move the camera quickly on Marble’s webpage and then stop, you can see the rendering gradually improve). The compression that spark.js applies to the Gaussian data may also reduce memory-access latency when rendering Gaussians. Performance optimization is on my TODO list, but it will take some time because I’m also busy with other projects.

[Arthur-Aillet]

I think the spark.js compression is a very good lead for optimization improvement.
I will try to implement it and create a pull request if the results I get are relevant.