Text-to-Video
Also known as: AI video generation, text to video, generative video, AI video creation
Text-to-video is a generative AI capability that creates video clips from natural language descriptions.
Like text-to-image, the user provides a prompt ('a slow-motion shot of rain falling on a lake at dusk') and the model generates a short video matching the description, including realistic motion, lighting, and in some cases sound. Text-to-video models are significantly more computationally demanding than image models, and the field advanced rapidly from 2023 with models such as OpenAI's Sora, Runway's Gen series, Kling, and Google's Veo.
Text-to-Video, explained at three levels
At home
Text-to-video AI takes a written description and produces a short video from it. If you type 'a golden retriever running through a field of sunflowers', the AI generates a clip of that scene, including realistic movement, lighting changes, and depth. The video is created from scratch; the AI is not searching for existing footage. The clips typically run from a few seconds to about a minute.
At work
Text-to-video tools are beginning to reach creative professionals as a way to prototype video concepts, create B-roll footage, or generate visual content without filming. Advertising agencies, content creators, and film pre-production teams are using them to visualize concepts quickly before committing to expensive production. Like text-to-image, commercial licensing varies significantly between providers.
At home
Text-to-video models face additional technical challenges compared to image models: they must maintain visual consistency of subjects, lighting, and scene across hundreds of frames; they must generate physically plausible motion; and they must handle camera movement realistically. The most advanced models use spatial and temporal attention mechanisms to track how elements in the scene should move relative to each other over time.
At work
Verification of video authenticity is a growing concern as text-to-video quality improves. Generated video clips can show events that never occurred, depict people saying things they never said, and appear indistinguishable from authentic footage to an untrained eye. Provenance standards (embedding metadata about whether a video was AI-generated) and video forensics tools are in active development. The Coalition for Content Provenance and Authenticity (C2PA) standard addresses AI-generated media labelling.
At home
Text-to-video generation is significantly more computationally expensive than image generation, requiring larger models and more processing time. A single high-quality video clip may require minutes of compute on high-end GPUs. This limits text-to-video to cloud API access for most users and makes on-premises deployment impractical for most organizations. Cost per second of generated video is orders of magnitude higher than cost per image.
At work
From a practitioner standpoint, text-to-video is currently most useful for concept visualization and creative ideation rather than final production. The technology's limitations: inconsistent hands, faces, and text in video; limited clip length; difficulty maintaining subject identity across a long sequence: mean that human-produced video still outperforms AI for most professional applications. The field is advancing extremely rapidly, and capabilities that seem research-only today may be production-ready within 12-18 months.
Real-world examples
At home
Jerome wanted to illustrate a short story he had written for a social media post. Rather than hiring a videographer, he used Runway to generate four short atmospheric clips matching scenes from his story, assembled them in a simple editor, and added his narration over the top.
At work
An advertising agency's creative team used Sora to generate concept video footage for a client pitch, visualizing three different visual directions for a campaign before any production budget was committed. The client approved one direction, and actual filming proceeded with a clear reference for what the final video should look and feel like.
Common confusions
- Text-to-video does not produce broadcast-quality video. Current models produce clips that look realistic under casual viewing but show artefacts under close inspection: unnatural motion, inconsistent details between frames, and difficulty generating accurate text or hands. They are tools for prototyping and ideation, not replacements for professional video production.
- Generated video is not archival footage. A video clip produced by a text-to-video model depicts events that never happened, in locations that may not exist. Using generated video as evidence, journalism, or in contexts where authenticity matters is deeply problematic and may constitute misinformation.
- Text-to-video is not the same as AI video editing. AI video editing tools modify existing footage (removing backgrounds, changing colours, upscaling resolution, removing objects). Text-to-video creates footage from scratch from a written description. The two technologies have different use cases and different output characteristics.
For analysts
In BA artefacts, text-to-video surfaces as a feature requirement in content creation platforms, marketing tools, and media production workflows. Key requirements: content policy controls (the system must prevent generation of non-consensual deepfakes, explicit content, and content depicting real people without appropriate consent), provenance metadata (AI-generated video must be labelled with generation metadata per C2PA or equivalent), and output quality controls (what is the acceptable clip length, resolution, and format?). Useful requirement: 'The platform must embed C2PA-compliant provenance metadata in all AI-generated video clips indicating generation date, model used, and original prompt.'
History & origin
Early text-to-video research produced short, low-resolution clips with inconsistent motion. The field advanced through 2022 and 2023 as diffusion model techniques from image generation were extended to video. OpenAI's demonstration of Sora in February 2024 (which showed 60-second photorealistic clips with consistent physics and camera movement) marked a significant public milestone. Google's Veo 3 (2025) added generated sound to its clips, and OpenAI announced in March 2026 that it was shutting down its Sora app and API. The models remain computationally expensive to run and are primarily accessed via API or subscription rather than running locally.