MORAIDICTIONNAIRE IA
Génératif

Text-to-Video

AI models that turn a single sentence into an animated video clip, frame by frame.

Imagine describing a scene in one sentence — "an astronaut riding a horse across Mars at sunset" — and getting a few seconds of coherent footage. That is the promise of text-to-video: a family of generative models that turn a natural-language description (the prompt) into a plausible sequence of animated frames, with no camera or actor.

How it works

Most recent systems rely on diffusion models, extended from images to video. The idea: a network learns to progressively denoise an initially random signal until it forms a video, conditioning this process on the text (via an encoder such as CLIP).

The main challenge is not producing a single beautiful frame, but ensuring temporal consistency: an object must not change color or vanish between frames. Modern architectures (often Diffusion Transformers) treat the video as a sequence of spatio-temporal patches and learn motion across time.

The iterative denoising step can be written, in simplified form, as:

$$x_{t-1} = \frac{1}{\sqrt{\alpha_t}}\left(x_t - \frac{1-\alpha_t}{\sqrt{1-\bar{\alpha}t}}\,\epsilon\theta(x_t, t, c)\right)$$

where $\epsilon_\theta$ is the predicted noise, conditioned on the text $c$.

Uses and limits

Strength Limit
Fast scene prototyping Short duration (a few seconds)
Lower production cost Imperfect physical consistency
Storyboards, ads, education Risk of deepfakes

Text-to-video collapses the gap between idea and moving image — but shifts the difficulty toward mastering the prompt and verifying what we are actually watching.

Explore the full AI dictionary →