3D generation

In short

3D generation is AI software that creates a three-dimensional object or scene from text, an image, sketches or other references. The result can be viewed from different angles and may include geometry, colour and texture.

Also called: Generative 3D, text-to-3D, 3D asset generation
Topic: Images, video, voice and music
Level: Some background helps
First described: Poole, Jain, Barron and Mildenhall introduced DreamFusion, a widely cited text-to-3D method, in September 2022.

A 3D generator interprets a prompt or reference and estimates the shape and appearance of an object. The output may be a polygon mesh, a neural scene representation or another format used for rendering. It is different from a flat image because the result is intended to support viewpoints beyond the input picture.

DreamFusion showed one route: it optimised a neural radiance field, or NeRF, whose rendered views were repeatedly judged by a pretrained text-to-image diffusion model. The paper's score-distillation method allowed text-to-3D generation without a 3D training set¹.

Newer systems can separate shape from texture. Hunyuan3D 2.0 describes a shape-generation model and a texture-synthesis model for high-resolution textured assets, with an image used as a condition². Generated geometry can still have holes, extra parts, poor topology or textures that fail when the object is animated or printed.

The technology appears in games, visual effects, product mock-ups, virtual spaces and 3D-printing workflows. A designer may use it to make a first-pass prop, then clean the mesh and materials by hand before putting it into a real project.

Example

A game designer prompts for a low-poly red camping lantern, receives a 3D asset, fixes its mesh in Blender and imports it into a test scene.

Why it matters

Making usable 3D assets is slower than making single images because shape, surface and every viewpoint must agree. AI can provide a starting point, but professional use still depends on checking and repairing the asset.

Questions people ask

What can 3D generation make?

It can make objects or scenes represented as geometry, rendered neural fields or textured assets. Systems such as Hunyuan3D 2.0 specifically target high-resolution textured 3D assets².

Is 3D generation the same as 3D scanning?

No. Scanning reconstructs an existing real object or place from measurements or images. 3D generation creates a new result from a prompt or reference, although some tools combine generation with reconstruction.

Can I 3D-print an AI-generated model?

Sometimes, but not automatically. The file needs suitable geometry, a watertight surface and a format accepted by the printing workflow. AI output often needs repair before it is ready for printing or animation.

Related terms

Tools that use it

Sources

Back to the glossary