AI-Generated 3D Assets with Blender MCP

I've been experimenting with using AI to create 3D assets for a hobby WebXR VR game, and thought it was interesting enough to share.

I used GPT to generate the concept art, then experimented with Opus 5.5 and Blender MCP to turn it into a 3D model. My first few attempts weren't great, but switching to an approach based on smaller building blocks produced a result that was much closer to the original than I expected.

I'm sure there are better ways to do this, but I ended up with a GLB that seems to perform well in the browser.

Here's the process, the iterations and a few things I learned along the way.

Starting point: a few GPT image prompts

Concept art plus a front / side / back sheet to use as the modelling reference.

GPT concept art of a glowing crystal stone arch, with front, side and back reference views

First prompt

Scripted stone blocks with glowing seams. Right vibe, but boxy: straight legs, a half-circle top and painted-on cracks.

First Blender attempt: a boxy stone arch with glowing seams

Second prompt

"The cracks don't feel real enough, add some." My verdict: "Overall this feels worse."

Second Blender attempt: the arch with heavier cracks

Third prompt

"It needs to be more like the source material." Silhouettes traced from the image and the art projected on. Near-perfect from the front, but broken from the top and inside, 180k triangles and not exportable.

Third Blender attempt: concept art projected onto traced silhouettes

A new approach

I decided to start from scratch with a GPT-generated view of separate rocks and crystals, with a plan to arrange them into the arch.

GPT-generated sheet of separate rocks and crystals

Building blocks

Opus generated nice-looking building blocks, with a few tweaks to remove the cartoon feel of the crystals and improve the textures.

The rock and crystal building blocks modelled in Blender

Final result

I was sceptical, but pleasantly surprised: when I asked it to arrange the building blocks to match the source image, the result was much closer than I expected.

The final arch assembled from the building blocks in Blender

What I learned

Bear in mind, this is a hobby project, and I'm sure there's better advice out there. But, for me:

Just dropping source images into Opus and asking it to build something wasn't enough.
I found that prompting the model to build smaller building blocks was much more productive.
Ultimately, it built something beyond my skill set and allowed me to generate a GLB file for a WebXR game, with a sensible triangle count and baked-in lighting and textures. It seems to perform really well in my tests.
Keep pushing yourself 😉

If you've experimented with this kind of Blender MCP workflow, especially if you've had better results, I'd love to hear any pointers or tips!

Originally posted on LinkedIn.