upvote
LLMs are a PCG system. This is one of my pet peeves, I've been working on a procedural generation constraint system for quite a while (way before LLMs) as a hobby, that came with a domain specific language, where you could specify constraints.

Things like the red key must be reachable before the red door, and there should be no path that bypasses the red door, etc. written as a set of constraints.

I was working on is because I realized, constraints tend to be similar in nature (graph cut, minimum/maximum shortest path etc.) algorithmically, but its usually quite hard to control designer intent.

The reason I shelved the project, was I asked an LLM a couple years ago, to generate a dungeon room graph, following a similar set of constraints, but stated in plain English, and it did a perfect job.

Which brings me to my final point. LLMs are amazing and flexible constraint solvers, and are woefully underutilized in scenarios like this.

reply
The trouble is the LLMs non-determinism and the "You're absolutely right, I placed the red key behind the red door rather than in front of it. This was load bearing for the whole dungeon and I blew it" moments.

With a hand authored constraint system if the red key is in the wrong place you have a bug in the constraint system that can be deterministically fixed, with a LLM you have to roll the dice on prompting again and go through expensive verification.

reply
Very humorous and accurate contemporary AI slop quote, thank you
reply
That’s exactly how I’ve used tools like Tripo and Meshy.

Use any image generation tool (or any image) and have a multimodal LLM like Gemini extract individual parts of the image “isolate with a transparent background” - then you can use those images for image-to-3d in those tools.

Don’t forget to use low or “smart” poly features otherwise you get too many vertices to the point you can’t performantly raycast etc.

But yeah - it’s there.

reply
> Don’t forget to use low or “smart” poly features otherwise you get too many vertices to the point you can’t performantly raycast etc.

I've found it's actually best to do maximal quality generation and decimate/meshopt to desired budget.

If you're raycasting against your raw art you're doing it wrong anyway, none of the big players do this precisely for the reason you mentioned. And beyond physics meshes, if you're doing anything where performance is a problem you probably want multiple LODs so you have to do this work anyway.

These days you can just ask your agent to build you a LOD pipeline and forget about it rather than sacrificing the art; Claude knows Blender.

reply
Never had success decimating in Blender. It’s WAY too many vertices. Unusably huge.

Decimating ends up ruining the look of the model to get it down to a game ready size.

Even a simple house model, or humanoid, using Decimate in Blender there is no way that will be even recognizable.

reply
Have you tried meshopt? It depends a lot on the art, view distance, and tuning, but it works well for my use cases.

(I agree with you about Blender though, I always hated Blender's simplify)

reply
Have you tried meshlab [1]? I've had pretty fantastic results reducing meshes from millions to some thousands.

[1] https://www.meshlab.net

reply
Side tangent: Smart Poly is often a paid feature, but you used to be able to steal it for free. Because in order to preview it, you download the GLB in the Network tab.

Some 3D AI providers still leak like that, but Meshy came up with their own binary format that splits the stream up into separate files that are assembled with a response it only gets if you have Premium.

Thought it was cool technology - and wonder if browser games in general are doing stuff like that to protect 3D assets from being easily ripped from the Network tab.

reply
> Still, think most people don't realize how good LLMs are at 3D these days, especially Fable 5

Can you explain more? Can I use Fable 5 to generate a 3d model from a couple of photographs?

reply