In plain words
You type "a red kite over a beach at sunset" and, a few seconds later, a picture appears. Nobody drew it and it is not a photo pulled from the internet. It was made fresh, just for those words.
These tools learned from hundreds of millions of pictures paired with their captions. From that, they learned which words go with which looks: what "sunset" does to the light, what a kite is shaped like, how sand looks next to sea. When you type a prompt, the model builds a picture that fits every part of it.
Well-known tools include Midjourney, DALL-E and Stable Diffusion. Most image models today work by starting from random noise and cleaning it up step by step, guided by your words. That method is called a diffusion model. Image making is part of Generative AI.
Everyday example
Describing a dream home to an architect
You tell an architect: "two floors, a big verandah, a neem tree by the gate." They have seen thousands of houses, so they sketch one that matches, even though it never existed. Change one word, "three floors", and the sketch changes. Text-to-image works like that sketching architect, at speed, for any scene.
Where the example breaks: The architect understands houses and asks questions when you are vague. The model just fills gaps with what is most typical in its training pictures, so it can get counts, hands, spelling inside the image or unusual combinations wrong without noticing.
How it actually works
A text-to-image system has two main parts: one that understands the words and one that paints the picture.
Turn the prompt into numbers
A text encoder turns your prompt into embeddings, lists of numbers that capture meaning. The original Stable Diffusion, for example, used the text encoder from OpenAI's CLIP, trained on image and caption pairs.
Start from noise
The image model begins with random noise, often in a compressed latent space rather than full pixels, which makes it much faster.
Clean up, guided by the words
Over many small steps, the model predicts and removes noise. At each step, a mechanism called cross-attention lets parts of the image look at the words, so "red" pulls colour into the kite and "sunset" warms the sky.
Decode to a picture
A decoder turns the final compressed result back into full-size pixels. Some newer systems instead generate images inside a large language model that handles text and images together.
- Early research
- 2015: alignDRAW, from the University of Toronto, drew small images from captions
- Went mainstream
- DALL-E (January 2021), then Midjourney and Stable Diffusion in 2022
- Open release
- Stable Diffusion's weights were released publicly in August 2022
- Where you meet it
- As of October 2026, built into chat apps, design tools and phone photo editors, as well as standalone apps
Questions about training on artists' work are covered in AI and copyright.
Questions people ask
How does AI generate images from text?
It learned links between words and looks from millions of captioned pictures. To make a new one, it usually starts from random noise and removes it step by step, checking against your words at each step, until a picture that fits appears. This is a diffusion model.
Does text-to-image AI copy existing pictures?
Normally it does not paste or collage stored images. It makes new pixels from learned patterns. But it can produce close copies of very famous images it saw many times, and it can imitate styles, which is why its use of artists' work is debated. See AI and copyright.
Why does AI get hands and text wrong?
Hands bend in many ways and are often partly hidden in photos, so the patterns are hard to learn. Text inside images needs exact letters, while the model works on overall looks. Newer models have improved a lot at both, but mistakes still happen.
What is a prompt in image generation?
The prompt is the description you type. Being specific helps: subject, setting, light, style and framing. "A red kite over a beach at sunset, wide shot, watercolour" gives the model more to work with than "kite".
How can you tell if an image was made by AI?
Look for odd details: hands, jewellery, background text, shadows that disagree. Some tools add a hidden watermark or Content Credentials you can check. None of these signs is proof on its own, so check the source too.