Text-to-video generation is the task of producing a video clip, typically a few seconds long, from a natural-language Prompt. It extends Text-to-image generation techniques into the time dimension, requiring a model to keep subjects, lighting, and physics consistent across frames rather than generating a single static image.
History
Early research systems in 2022, including Meta's Make-A-Video and Google's Imagen Video, adapted Diffusion model techniques from image generation but produced short, low-resolution, often incoherent clips. Runway's Gen-1 and Gen-2 models, released in 2023, were among the first widely accessible commercial tools. The field's most publicized moment came with OpenAI's Sora in February 2024, previewed as capable of generating minute-long, high-resolution clips and described by OpenAI as a step toward general-purpose "world simulators," though it was not released broadly until later. Google DeepMind's Veo followed a similar trajectory, with Veo 3 in 2025 adding synchronized audio generation and integration into YouTube's creative tools. Chinese labs moved quickly into the space as well: Kuaishou's Kling and ByteDance's Seedance both topped independent video-generation leaderboards during 2025, reflecting an increasingly global and competitive field.
Technical approach
Most text-to-video systems extend a latent diffusion model with temporal layers that model motion across frames, sometimes combined with a Transformer (architecture)-based architecture operating over spacetime patches rather than single images, an approach OpenAI described for Sora. Generating a coherent video is substantially harder than a single image because the model must maintain object permanence, plausible physics, and consistent lighting over time, and errors tend to compound across frames. Many systems also support image-to-video, animating a still image supplied by the user, and some allow control over camera movement or the extension of an existing clip.
Applications and reception
Text-to-video tools have been adopted for advertising, previsualization, social media content, and short-form entertainment, with some studios experimenting with the technology for pitch materials or background footage. The technology drew comparisons to earlier disruptions in text-to-image generation and AI music generation, including objections from actors, filmmakers, and unions over the use of copyrighted footage in training data and the potential displacement of production jobs; the 2023 Hollywood strikes cited AI-generated performance and likeness concerns as a central issue. Concerns about Deepfake misuse are especially acute for video, since a realistic fabricated clip of a real person can spread faster and be more persuasive than a still image, prompting several platforms to require or explore AI watermarking for AI-generated video.
Limitations
As of 2025, most publicly available systems remain limited to clips of a few seconds to roughly a minute, with cost and compute scaling steeply with resolution and length. Fine control over precise dialogue, complex multi-character interactions, and long-range narrative coherence remained open problems, an area of active competition among OpenAI, Google DeepMind, Runway, and ByteDance (AI).