text-to-image
/ TEKST-too-IM-ij /
Text-to-image is the technology that turns a written description into a brand-new picture: type "a red panda reading a newspaper in a cozy cafe, watercolor style," and out comes an image matching that description, one no one ever drew. It is the headline trick of tools like Stable Diffusion, Midjourney, and DALL·E, and it stitches together two pieces — a way to understand language and a way to produce images.
Under the hood, two things cooperate. First the system has to connect words to visual concepts, usually via a model trained on hundreds of millions of image-and-caption pairs scraped from the web, so it learns that the words "red panda" go with a certain furry shape and color. Then an image generator — almost always a diffusion model — uses that understanding to steer its denoising, pulling random static toward a picture that fits the prompt. The text acts as a guide rail at every step of generation.
Be candid about the limits and the stakes. These systems excel at look and mood but stumble on anything requiring precise structure or logic: counting ("exactly five apples"), spelling words in the image, correct anatomy, or faithful spatial relations ("the cup to the left of the book"). They also raise real, unresolved questions — the training images often included copyrighted and personal photos scraped without consent, and the same tech enables convincing fakes of real people. The pictures are impressive; the ethics and the reliability are genuinely unsettled.
Prompt: "an astronaut riding a horse on the moon, photorealistic." No such photo exists or could exist, yet the system produces a convincing one — because it learned separately what astronauts, horses, and the moon look like, and the diffusion steps blend them into a single coherent scene.
Combining learned concepts into a scene that never existed.
Text-to-image models are weak at exactly the things that need precision: counting objects, spelling text inside the picture, correct hands, and spatial relations. They also carry unresolved copyright and consent questions, since much of their training imagery was scraped without permission.