Really cool we’re at the point where you can build custom, personalized tools for your own niche 3D workflows. Built this with GPT-5.5, Spark, and Marble yesterday to get better precision + control for creating collider meshes for 3DGS experiences. Live:

Dude this is sick

Thanks! 🙏 my workflow was initially blender (which I love) but no native support for 3DGS. And any of the plugins if you bring a splat over 1 mil in size the fps tanks 👀

@carrabre That's pretty damn cool, do you plan to release this?

@carrabre Yes! You can try it now here!

@carrabre Nicely done - open source this sucker 😎

@carrabre 100% will soon! @sparkjsdev is the renderer I use which is open source!

Is the flow be like Generate 360 degree pano from gpt5.5 and then to marble.?

Yes! I’ve done that but also use Midjourney too

Amazing Ian! 😮😍

Thank you Matt!!

for one second i thought you're moving things around and i was like NO WAY!!!

Sharper collider meshes open new dimensions in 3D.

The real unlock isn't the tools, it's that domain experts can finally build for themselves without a middleman. A generalist dev would've shipped something 80% wrong for your workflow. What specific part of collider mesh creation was impossible to get right with off-the-shelf solutions?

totally glitchy nothing works baby. thats how all these tools are the demos look great but when you use it it looks forible.

Oh, this is awesome. Any thoughts on automated collider generation for world-model-generated scenes? We’re looking into this exact thing right now with Marble for our full-on AI 3D gaming engine.

are these individual marble scenes? if so there is a collider mesh export per scene. also in api.

h nice, let me try using them in a quick prototype. Haven’t gotten to this part yet. Why did you build your own if they already exist though?

good q! For more complex scenes sometimes manual collide can work better if you are not trying to scale it. For example I'm just building out a game or two, for a entire engine might be better to scale it. By doing it manually the collider mesh is pretty much near perfect. super low poly count, very optimized and ready to ship. Also at this moment do not have a solution for stitched or larger scenes so I also use this for now. But we are working hard to automate a lot of this! If you have any input on how to make it better please let me know 🙏

Perfect! Then I guess we can just wait a bit for those 🙂 One feature request you’re probably already aware of is a WebGPU renderer (Spark currently runs on WebGL). We’re building a native engine and exploring a switch to WebGPU for rendering - you can compile it to native targets as well. That would give us a single, consistent render pipeline and let us reuse Spark outside of three.js ecosystem in our custom environment without rewriting Spark shaders etc.

Love it :) is the entire primitive the collider or just the splats it highlights?

ohhh only the primitive. But that would be really cool if some logic to get the splat only part 👀

The real unlock isn't the tools, it's that domain experts can finally build for themselves without a middleman. A generalist dev would've shipped something 80% wrong for your workflow. What specific part of collider mesh creation was impossible to get right with off-the-shelf solutions?

using a git link with claude code is definitely the fastest way to get a repo digested but the real headache starts when you want that 3dgs precision logic to run on a schedule without babying it.

you made your own sparkedit thats pretty cool

collider mesh precision is where most ai tools fall apart. convex decomposition decisions need to be deterministic or your physics sim gives different results every run.

AI people conflate the tool with the output. "My pen drew a line!" So?

Niche tools push 3D workflows into artful clarity.

@drfeifei Cool ! How to make it?