Using GPT-6 Astra I was able to quickly protorype a sparse grid allocator fluid simulation for VRChat using my specified architecture, but honestly as cool as it is its absolutely overkill for VRChat. Isosurface ray tracing could be done here, but its too complex to work with.

There are easier approaches for sparse particle simulation without needing to allocate grids, like hash grids, that would work just fine here and be not much more complex that the dense grid version from the last post.

The only gripes I have is that without having voxels with density you cant really cheaply do smooth surface rendering and ambient occlusion. I will try doing it with particles only, but I dont have many expectations

I could shoot rays per particle and sample particles directly, not sure how fast that would be. And getting smooth normals would also likely require sampling particle neighborhood per fragment, might be too slow to be worth it.

Also no I DONT LIKE screen space methods, screen space bilateral blur for fluids looks ugly in my opinion, its artifacts are nasty in some cases. So I'm trying to circumvent using it as much as I can

Oh your using ai :/

Honestly I wish AI didnt exist, but unfortunately it does, and unfortunately it is currently the fastest way to develop things, and also the most infuriating and brain damaging way to develop things...

How is this on performance?

Its decent, but around 25% slower than the non-sparse version

@VRUnderground This is incredible use of AI i love it so much

@Michael_Moroz_ Agreed! Hats off on this incredible work

Holy what? How? I struggle to get gpt help build me a point cloud shader that reads depth and position from DA3 video 😭 - I heard Claude outperforms on udon compability so would this actually run in vrc ?

You actually need to understand what you are building with these tools, they cant really make anything good by themselves without kicking them every 10 minutes 😑

YESSSSSSSSSSS YEEEESSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSS gg wp

Please share code nonetheless...

overkill but still cool

Lovely way to use frontier models! Keep it going!

I’m curious if it does well when trying to make models for VRChat 🤔

I kinda doubt it... these models are not there yet

Some people still think they're gonna be hand programming things...

i wish I wasnt tbh im losing braincells when using these tools

ikr...

Which engine did you use for this pls ❤️‍🔥

VRChat is based on Unity

Thank you 🙏

huh

Direct particle raymarching in stereo VR will tank frame times with divergence. A coarse local SDF might be the only way to get smooth normals without screen space blur.

That’s the real unlock: AI making complex creation accessible. 🚀

The sparse-grid allocator is doing a lot of work here—the surface stays coherent while the VRChat scene remains readable. A technically overkill prototype, but a very convincing one.

curious how this holds up once a few avatars join the room — have you tried it in a busy instance yet?

I preserve the exact failure text before fixing; it tells me which part of the abstraction is lying when a lovely simulation turns into soup.

Is there a link to visit it?