Images, video, voice and music
Text-to-video
In short
Text-to-video is an AI method that creates a moving-image clip from a written description. The model predicts visual content across successive frames so the scene, movement and camera view fit the prompt as well as possible.
- Also called
- Text-to-video generation, AI video generator, T2V
- Topic
- Images, video, voice and music
- Level
- Beginner
- First described
- Vondrick and colleagues published Video Generation From Text in 2017, an early paper explicitly studying video generation from text descriptions1.
A text-to-video system takes a prompt describing a scene, action, subject or style and generates a sequence of frames. Compared with text-to-image, it must also keep objects, motion, lighting and camera changes reasonably consistent over time12.
Many systems use diffusion models. A text encoder represents the prompt, a base model generates a low-resolution or short video, and extra models can improve spatial detail and the timing or number of frames. Google's Imagen Video paper describes this as a cascade of video diffusion and spatial and temporal super-resolution models2.
Text-to-video does not record a real event or understand a scene like a human camera operator. It generates a plausible sequence and can create errors such as changing faces, impossible physics, unstable text or objects appearing and disappearing. Prompt adherence and realism therefore need to be checked, especially for news, evidence or safety-critical communication13.
The method is used for storyboards, adverts, visual effects, social clips, education and creative experiments. OpenAI's Sora report described a model that could generate videos up to a minute from text, while the company later said the Sora product was no longer available as of 26 April 20263.
Early research also used paired text and video data to train a model to generate clips from descriptions. Later work reduced dependence on such paired data by adapting knowledge from image-generation systems14.
Example
A teacher making a lesson about weather could request a short clip of clouds gathering over a city before rain, then review it for mistakes before showing it to students.
Why it matters
Text-to-video makes moving visual drafts easier to produce, but a convincing clip can still be synthetic, inaccurate or misleading.
Questions people ask
How does text-to-video work?
The system encodes the prompt and generates a sequence of video representations, often with diffusion and separate stages for spatial detail and temporal consistency. The exact architecture varies by model24.
Is text-to-video the same as animation?
No. Animation is a wider creative and production practice. Text-to-video is one AI-based way to generate moving images, and its output may still need editing, sound, compositing or human animation.
Can text-to-video make realistic footage?
It can produce clips that look realistic, but realism is not proof that an event happened. Models can introduce visual or physical errors, so generated footage should be labelled and important claims independently checked3.
Is Sora still available?
OpenAI's Sora page says the Sora product was no longer available as of 26 April 2026. Availability can change, so check the provider's current notice3.
Related terms
Tools that use it
- RunwayNew York AI video company whose phone app and website turn a typed idea or a photo into short realistic clips, using its own Gen-4.5 model and rivals' models.
- VEEDA browser video editor that trims clips, adds subtitles, translates speech and can generate short videos from typed prompts.
- ViduMakes short video clips with sound from a written description, a still image or reference pictures.
- Adobe FireflyAdobe's image and video generator, trained on licensed stock, with Google and OpenAI models bolted on.