MORAIDICTIONNAIRE IA
Génératif

Text-to-Image

Turning a sentence into a picture: how AI paints from words.

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.

Explore the full AI dictionary →