Log in

AI for Three.js

Text to 3D: Generating a 3D Model From a Prompt

Text-to-3D AI tools generate a 3D model directly from a written description, such as "a low-poly pine tree" or "a worn leather armchair," without any reference image. That makes them fastest for concepting, placeholder assets and stylized objects, but generally less geometrically accurate than image-to-3D when you need to match a specific real object.

Last updated . Verified against three.js r181.

Read the FAQJump to code
Isometric diagram of AI skills, tools and workflows feeding a Three.js scene

[01]

When text-to-3D is the right tool

Text-to-3D shines specifically when there's no reference photo to work from at all: you're concepting something that doesn't exist yet, blocking out placeholder assets for a scene before final art is ready, or generating a stylized/non-photorealistic object where "close enough and fast" beats precise geometric accuracy. It's a worse fit when you need to match a specific real-world product or object exactly. That's image-to-3D's job.

[02]

Prompting for a model you can actually use

The prompts that produce cleaner, more usable geometry share a few traits: describing the object's overall silhouette and material clearly ("a rounded ceramic mug with a thick handle" beats "a mug"), specifying style explicitly if you don't want photorealism ("low-poly," "stylized," "hand-painted texture"), and keeping the prompt to one clear subject. Compound prompts describing a whole scene tend to produce a single confused blob rather than distinct, usable objects.

Iteration matters more than getting the prompt right first try. Most tools let you regenerate quickly, and two or three attempts with small prompt adjustments usually beats over-engineering one prompt.

[03]

From generated model to a Three.js scene

Once generated, the workflow converges with any other AI-generated asset: check poly count, verify texture resolution, confirm the export format is glTF/GLB, and normalize scale on import since text-to-3D has even less real-world size grounding than image-to-3D. See the full AI 3D model generators guide for the complete cleanup checklist before loading with GLTFLoader.

Code

Loading and scaling a text-to-3D model (r181)
1import { GLTFLoader } from 'three/examples/jsm/loaders/GLTFLoader.js'
2
3const loader = new GLTFLoader()
4loader.load('/models/generated-from-prompt.glb', (gltf) => {
5  // Text-to-3D output has no real-world scale reference at all. Always
6  // normalize explicitly rather than trusting the file's stated units.
7  const box = new THREE.Box3().setFromObject(gltf.scene)
8  const size = box.getSize(new THREE.Vector3()).length()
9  gltf.scene.scale.setScalar(2 / size) // 2 units tall, adjust to your scene
10  scene.add(gltf.scene)
11})
Stylizing a generated model toward "low-poly" with TSL
1import { texture, positionLocal, normalLocal, dot, vec3 } from 'three/tsl'
2
3// A cheap flat-shaded look, useful for smoothing over a generated model's
4// imperfect detail rather than fighting it: recompute shading per-face
5// instead of relying on the model's (often noisy) vertex normals.
6const material = new THREE.MeshBasicNodeMaterial()
7const lightDir = vec3(0.5, 0.8, 0.3).normalize()
8const ndl = dot(normalLocal, lightDir).max(0.2)
9material.colorNode = texture(generatedAlbedo).mul(ndl)

Learn this properly

Learn Practical TSL

Your First Node Material

Useful for re-shading a generated model once it's in your scene.

Start the lesson (6 minutes) →

F.A.Q

Frequently asked questions

The questions that come up most often on this topic.

Is text-to-3D or image-to-3D more accurate?

Image-to-3D generally produces geometrically more accurate results when you're trying to match a specific real object, since it has visual reference. Text-to-3D is faster and better suited to concepting, placeholders and stylized objects where exact accuracy matters less.

How do I write a good prompt for text-to-3D generation?

Describe the object's silhouette and material clearly, specify a style if you don't want default photorealism, and keep the prompt to one clear subject rather than a whole scene. Compound prompts tend to produce confused, unusable geometry.

Can text-to-3D models be used commercially?

Depends on the specific tool's terms. Check the license for commercial use before shipping generated output in a paid product, as terms vary between tools and can change.

Building this for a client?

I take this kind of work as a white-label subcontractor for agencies: scoped in writing, delivered under your brand, documented for your team.

For agencies →

Keep reading

Image to 3D

Image-to-3D AI tools generate a 3D mesh from one or more reference photos. Feed in a photo of an object, get back a text…

AI 3D Model Generators

AI 3D model generators turn a text prompt or a reference image into a usable 3D mesh. Tools like Meshy, Tripo and Rodin …

AI Game Generators

AI game makers and generators are tools that use AI to scaffold or generate parts of a game: level layouts, sprites, sim…

All AI for Three.js

Back to the ai hub.