How does GPT-6 actually generate 3D models in Release Demo? Via Codex + Local Blender or MCPs/APIs?

Codex
1

Hi everyone! I’m trying to understand how the GPT-6 demo generates home design 3d in Blender / Unreal Engine, and how we could reproduce a similar workflow.

My current guesses are:

  1. GPT-6 → Blender/UE API → .glb/.blend
  2. GPT-6 → Codex/Computer Use → controls local Blender
  3. GPT-6 directly generates a 3D model/mesh (I doubt this, since GPT-6 is still primarily a language/reasoning model)
  4. GPT-6 → spatial/scene description → Python/bpy → Blender → .glb

For example, GPT-6 could generate something like with parts, geometry and location:

Floor: 3m × 8m @ (0,0,0)
Wall: 3m high @ (0,4,1.5)
Lamp: 1m × 2m × 2m @ (0,5,2)

Then a python script converts that text description into Blender geometry.

So what is actually happening in the demo?

If it’s mainly #1 or #2, does that mean the impressive result comes primarily from better reasoning + coding + tool use, rather than a dedicated spatial ability improvement?

Or does GPT-6 actually maintain some form of 3D/spatial representation while planning the scene, with Blender/UE simply executing and rendering the result?

I’d especially love to hear from anyone who has reproduced the demo using MCP, Codex, Blender Python/bpy, or 3D-generation APIs and share your thoughts on the roadmaps?

2

It runs Blender in headless mode from the command line.

An embedded Python script controls Blender through its bpy API, using calls such as:

bpy.ops.mesh.primitive_cube_add(...)
bpy.ops.wm.save_as_mainfile(...)
bpy.ops.render.render(write_still=True)

There are other ways too, like with MCP but the scripts only require that you have blender installed, and astra “knows” how to use it.

You can just install Blender, and ask codex to use it to help render the objects.

3

So a bit like its headless Chrome use. Very cool.

How big is a blender install? (I looked: 400MB, 1GB free recommended ?)

4

Hi aprendendo.next,
Thanks! Actually, I’ve been experimenting with the Blender bpy method of a Lego Build example by asking GPT to write Python code to generate LEGO part-building scripts. And the demo is pre sure from manipulating the bny API rather then directly generate 3d model blueprints. I am comparing the “GPT + Codex Python code + Blender bpy API and render methods” with the “GPT Models Directly Infer 3D parts blueprint + Direct Render in Viewer” methods, and want to find out is it the real spatial intelligence?

I can say that GPT 6 model are capable of writing good enough code to generate very complicated 3D structures by bpy python APIs. However, directly generating a blueprint file and inferring the precise centroids and positions of individual parts as output of the LLM is still quite hard for models to complete and challenging.

Method 1 or 2: ChatGPT/Codex + Python Code + Blender Python API

Prompt: Generate 3D Premium white and blue double-deck yacht Lego Builds with studs parts inventory. Write Python Code to Call Blender Bny
Write a python code generate_lego.py and output glb file.


## some demo generation code of contour, location of the parts 
def create_cube(name,location,dimensions,material,bevel=0.025):

    bpy.ops.mesh.primitive_cube_add(
        location=location
    )

    obj = bpy.context.object

    obj.name = name
    obj.dimensions = dimensions
    bpy.ops.object.transform_apply(
        location=False,
        rotation=False,
        scale=True
    )
    if bevel > 0:
        modifier = obj.modifiers.new(
            "LEGO Soft Edges",
            "BEVEL"
        )
        modifier.width = bevel
        modifier.segments = 2
    obj.data.materials.append(material)
    return obj

Pasted Graphic 16
Pasted Graphic 161920×765 153 KB

Method 3 or 4: GPT Models Directly Infer 3D parts blueprint + Direct Render File

Prompt: Generate 3D Premium white and blue double-deck yacht Lego Builds with parts inventory. Generate Parts Inventory Description that works Craftsman Agent Lego Builds APIs Outputs Blueprint with
color size and centroid position (x,y,z) formats

[{"color":"bright_blue","size":[1,1,1],"position":[4,0,0]},{"color":"bright_blue","size":[1,1,1],"position":[4,1,0]},{"color":"bright_blue","size":[1,1,1],"position":[4,2,0]},{"color":"bright_blue","size":[1,1,1],"position":[3,3,0]},{"color":"bright_blue","size":[1,1,1],"position":[4,3,0]},{"color":"bright_blue","size":[1,1,1],"position":[5,3,0]}
.... 
]

3D Model OBJ Viewer of Craftsman Agent of the Lego Blueprints : Lego Luxury Yacht Assembly Pack | Lego | Craftsman Agent

Final 3D Models Preview Image

Pasted Graphic 15
Pasted Graphic 152084×1740 341 KB

So even for GPT-6, building complex things still needs to manipulate of Blender APIs rather than direct generation of the 3d blueprints.