Text-to-image

In short

Text-to-image is an AI method that creates or edits a picture from a written description. The model turns the prompt into a representation of the requested scene, then generates pixels that aim to match it.

Also called: Text-to-image generation, AI image generator, T2I
Topic: Images, video, voice and music
Level: Beginner
First described: Reed and colleagues introduced a generative adversarial approach for synthesising images from text descriptions in their 2016 paper¹.

A text-to-image system takes a written prompt such as a description of a place, object or style and returns an image. It learns relationships between language and visual patterns from training examples, then uses those relationships to construct a new result¹².

Many newer systems use diffusion: they learn to reverse a controlled process that adds noise to images, starting from noise and repeatedly making the result more like an image that fits the text. Other designs generate compressed image tokens one after another, as the original DALL·E paper described²³.

The output is not a photograph retrieved from a catalogue simply because the prompt mentions it. It is a generated image, and the model can misread a request, invent details, struggle with text or produce inconsistent hands, objects or counts. A prompt is guidance, not a guarantee of factual accuracy or ownership²⁴.

Text-to-image tools show up in illustration, advertising, concept design, education, games and photo editing. The first widely cited research systems generated limited examples such as birds and flowers; later systems such as DALL·E and Imagen broadened the kinds of scenes they could attempt¹²³.

Example

Someone planning a birthday party can type 'a cheerful poster with yellow balloons, a blue cake and space for the date' to make a draft design, then edit the wording and layout themselves.

Why it matters

It lets people describe a visual idea in ordinary language, while also raising practical questions about errors, consent, copyright and how generated images are labelled.

Questions people ask

Is text-to-image the same as image search?

No. Image search retrieves or ranks existing images. Text-to-image generates a new image from learned relationships between text and visual patterns, although a product may offer both features¹².

How do text-to-image models work?

A text encoder represents the prompt, and a generator uses that information to produce an image. Diffusion systems usually begin with noise and remove it step by step; other systems generate image tokens sequentially²³.

Can text-to-image models make exact words?

They can attempt it, but generated lettering has often been less reliable than objects, colours or broad composition. Results depend on the model and prompt, so important wording should be checked and corrected.

Related terms

Tools that use it

Sources

Back to the glossary