Black October, explainedAppsGuidesAI DictionarySmall FixesOn RecordAboutFollow the lab

3D generation

Also called text-to-3D, AI 3D model

3D generation is AI that makes a 3D object or scene, one you can turn and view from any side, from words, a photo or a short video.

Images, voice and videoCurious

In plain words

A photo is flat. A 3D model is an object you can spin around on screen, light from any side, drop into a game, or send to a 3D printer. Making one by hand can take a skilled artist many hours.

AI now does it in two main ways. The first rebuilds something real: walk around an object filming it on your phone, and the AI works out its full shape from all those views. The second invents something new: type "a clay pot with blue flowers", or give one photo, and the AI imagines the sides it cannot see.

This is used for game and film props, online shopping where you rotate a product, virtual home tours, and quick prototypes before anything is built.

Everyday example

A potter shaping a matka from a description

Describe a matka to a potter: round, wide mouth, a band of blue around the middle. The potter shapes it so it looks right from every side, turning the wheel and checking. Text-to-3D works similarly: it keeps checking views from many angles and adjusts the shape until every view matches your words.

Where the example breaks: The potter has hands and real clay, and knows what the inside of a pot is for. AI only makes the shape look right from outside. Hidden parts, thin details and the backs of objects are often guessed badly, and printed objects may not be strong or even solid.

Go deeper

How it actually works

3D generation borrows heavily from 2D image AI, because there are far fewer 3D models to learn from than photos.

  1. Rebuild from many photos

    NeRF (2020) trains a small neural network to answer: at this point in space, seen from this direction, what colour and how solid is it? Given enough photos from different angles, it can render new views. 3D Gaussian splatting (2023) stores the scene as millions of soft coloured blobs and renders it in real time, at 30 frames a second or more at full HD.

  2. Invent from words using a 2D model

    DreamFusion (2022) made 3D objects from text without any 3D training data. It starts with a random 3D shape, renders it from a random angle, and asks a text-to-image diffusion model how to make that picture fit the prompt better. Repeating this from many angles shapes the object.

  3. Go faster with one photo

    Newer systems first make an image, or take yours, and then predict a 3D shape from it directly. OpenAI's Point-E (2022) made rough 3D point clouds in one to two minutes on a single GPU. Meta's SAM 3D (November 2025) rebuilds objects and human bodies from one everyday photo.

  4. Export to a usable format

    Results come out as a mesh (a surface of tiny triangles, the standard in games and printing), a point cloud, or Gaussian splats. DreamFusion's results, for example, could be converted to meshes and relit with any lighting.

Reconstruction milestone
NeRF, March 2020: new views of a scene from a sparse set of photos
Text-to-3D milestone
DreamFusion, September 2022, by researchers at Google Research and UC Berkeley, guided by Google's Imagen image model
Real-time rendering
3D Gaussian splatting, August 2023: 30 frames a second or more at 1080p
Uses
Game and film props, product views for online shops, room and building scans, quick prototypes for 3D printing

Understanding the 3D shape of real things also matters for physical AI, where robots must handle actual objects.

Questions people ask

Can AI turn a photo into a 3D model?

Yes, with caveats. From one photo, AI has to guess the sides it cannot see, so the back may be wrong. From many photos or a short video circling the object, it can rebuild the real shape much more accurately.

What is text-to-3D?

Text-to-3D means typing a description and getting a 3D object. Early methods such as DreamFusion shaped the object by repeatedly checking rendered views with a 2D image generator. Many newer systems first generate a picture and then lift it into 3D.

What is a NeRF?

NeRF stands for neural radiance field. It is a way of storing a scene inside a small neural network that, for any point and viewing direction, returns a colour and density. Rendering those along camera rays makes new photos from angles nobody shot.

Can I 3D print an AI-generated model?

Often, after cleanup. AI models may have holes, paper-thin walls or messy surfaces that printers cannot handle. Designers usually repair the mesh in 3D software first. For anything that must bear weight or fit exactly, a person should check it.

Further reading

Prepared by Mundhra Labs. Reviewed by Aditya Mundhra. Checked October 2026.