Diffusion neural networks for 3D graphics share a common flaw: they generate "baked" irregular meshes that require painful manual retopology.
The open 4B-parameter Qwen3.5-4B-GLB-3D-Coder model offers a different approach: procedural Three.js code with built-in export to a binary GLB container.
Example of a generated script:
import * as THREE from 'three';
import { GLTFExporter } from 'three/examples/jsm/exporters/GLTFExporter.js';
const scene = new THREE.Scene();
const mesh = new THREE.Mesh(
new THREE.CylinderGeometry(0.5, 0.8, 2, 16),
new THREE.MeshStandardMaterial({ color: 0x0088ff, roughness: 0.2 })
);
scene.add(mesh);
new GLTFExporter().parse(scene, (glb) => {
const link = document.createElement('a');
link.href = URL.createObjectURL(new Blob([glb], { type: 'application/octet-stream' }));
link.download = 'asset.glb';
link.click();
}, { binary: true });
Running inference via Ollama requires less than 9 GB of VRAM:
ollama run qwen3.5-4b-glb-3d-coder "Generate a lantern with GLB export"
The binary .glb can be dragged and dropped into Godot 4.x with materials preserved.
