Generative

Text to 3D

Generating a three-dimensional model from a written description.

#ai
#generative
#retopology
#quad-topology

Quick answer

What is text to 3D?

Generating a three-dimensional model from a written description. A generative AI model reads a prompt such as "a mossy stone well", produces the object's geometry and surface textures, and outputs a mesh that 3D software can open. It is used to prototype game assets, props and product concepts without modelling every shape by hand.

How text to 3D works

Early research systems did not train on 3D data at all. DreamFusion (2022) started from a random 3D representation, rendered it from many angles, and used a 2D image diffusion model to score whether each render matched the prompt, nudging the shape until every view agreed. It worked, but it was slow: each object needed its own lengthy optimisation run.

The next step was models trained directly on large collections of 3D objects. OpenAI's Shap-E (2023) generates the parameters of a 3D shape in one pass, which makes generation dramatically faster. Today's tools typically combine both lines of work: they generate consistent views of the object, then reconstruct a textured mesh from those views.

Whatever the internals, the output a user cares about is the same: a polygon mesh, texture maps and a file format that other software can read.

What the output is good for

Text to 3D is strongest wherever many decent assets beat one perfect one: props and set dressing for games, background objects in product renders, blockouts for level design, concept exploration before a modeller commits to a direction, and placeholder assets that keep a project moving.

It is weaker where an exact shape matters. A prompt cannot carry precise dimensions, and two generations from the same words can differ a lot. When you already know the silhouette you want, an image or sketch is a tighter specification than text.

What still needs a human

Raw generated meshes tend to be dense and irregular: the shape reads well in a render, but the surface is a mass of small triangles. That is fine for a static background prop and a problem for anything that will be edited, subdivided or animated, which is why retopology usually follows generation.

Texture quality, scale, pivot placement and naming also need checking before an asset goes into a production scene. Treat generated models the way you would treat an asset from a marketplace: inspect the wireframe, not only the render.

Text to 3D, image to 3D and sketch to 3D

The three differ only in how the object is described. Text gives the model the most freedom and the least control. A photo pins down appearance but only from one side. A sketch sits in between: it fixes the silhouette and proportions you care about while leaving materials open. Many tools, Vi3W included, accept all three, and switching input is often the fastest way to correct a result that keeps coming back wrong.

Text to 3D in Vi3W

Vi3W generates models from text prompts, sketches or photos in the browser, retopologises the result automatically, and exports glTF, FBX or OBJ. The text-to-3D page covers the workflow and its limits.

Frequently asked questions

Is there a 3D AI?

Yes. "3D AI" usually means a generative model that creates 3D models from text, images or sketches. These tools output real meshes with textures that you can rotate, edit and export to Blender, Unity or Unreal Engine, unlike image generators, which only produce flat pictures that look three-dimensional.

Can ChatGPT create a 3D model?

Not directly. ChatGPT is a language model, so it cannot generate a textured mesh on its own. It can write code that produces simple geometry, such as an OpenSCAD file or a Blender Python script, and it can write detailed prompts. For an actual textured model, pair it with, or use instead, a dedicated text-to-3D generator.

Is 3D AI free?

Often partly. Most text-to-3D tools offer a free tier with a limited number of generations, then charge for regular use, because generating 3D models runs on expensive GPUs. Some research models are open source and free to run if you have your own hardware. Check whether free outputs can be used commercially before you ship them.