Text-to-image is a family of generative AI models that produce a brand-new image from a plain-language description called a prompt. Picture a painter who has never seen your idea yet sketches it instantly from your words alone — that is exactly what systems like DALL·E, Midjourney and Stable Diffusion do.
Linking words to pixels
The core challenge is aligning two worlds: text and images. Models learn this alignment from massive datasets of (image, caption) pairs. A text encoder (often inspired by CLIP) turns the prompt into a meaning vector that steers visual generation. The model does not "copy" existing images; it samples a fresh one from the learned space.
The diffusion mechanism
Most modern systems rely on diffusion. The idea: learn to denoise. During training, Gaussian noise is gradually added to an image, and the network learns to reverse the process. At generation time, it starts from pure noise and denoises step by step, conditioned on the text.
$$x_{t-1} = \frac{1}{\sqrt{\alpha_t}}\left(x_t - \frac{1-\alpha_t}{\sqrt{1-\bar{\alpha}t}}\,\epsilon\theta(x_t, t, c)\right)$$
where $c$ is the text conditioning and $\epsilon_\theta$ the predicted noise.
Families of approaches
| Approach | Principle | Examples |
|---|---|---|
| Diffusion | Iterative denoising | Stable Diffusion, Imagen |
| Autoregressive | Sequential pixels/tokens | DALL·E (first version) |
| GAN | Generator vs discriminator | StyleGAN-T |
Text-to-image makes visual creation as easy as writing a sentence — yet it raises pressing questions of copyright, bias and misinformation.