text-to-video model

A generative modelgenerative model, n.: A model that learns what a kind of data looks like — faces, sentences, songs — well enough to produce new examples that resemble the originals without copying them. that produces video conditioned on a text prompt and, usually, on images (image-to-video) or other video. Most current systems are latent diffusionlatent diffusion model, n.: A diffusion image generator that works on a shrunken, compressed version of each picture instead of the full-size pixels, which makes it far cheaper to train and run. or flow matchingflow matching, n.: A way to train generators that learns, for every point between random noise and real data, which direction to move, so that noise can be steered smoothly into new examples. models: a video autoencoder compresses frames in space and time, and a diffusion transformerdiffusion transformer, n.: A diffusion image or video generator whose core network is a transformer — the architecture behind chatbots — instead of the older image-specific design, which lets it improve more predictably as it grows. denoises 'spacetime patches' of the latent, attending across frames so that objects, lighting and motion stay consistent; some models instead generate frames or tokens autoregressively. Training uses very large collections of captioned video, often recaptioned by vision-language models, together with images. Since 2025 leading models also generate synchronized audio — speech, sound effects and music. Clips typically last 5–30 seconds at up to 1080p or 4K, with tools for extension, camera control and character consistency; physics errors, lapses in object permanence and drift over long durations remain common. Prominent systems as of October 2026 include Google's VeoVeo, prop. n.: Google's video-generating AI, which since 2025 creates short clips complete with matching sound and dialogue from a written prompt. and Gemini Omni, ByteDance's Seedance, Kuaishou's Kling, Runway's Gen series and Alibaba's Wan, several versions of which have open weights; OpenAI discontinued SoraSora, prop. n.: OpenAI's video-generating AI, unveiled in 2024 with startlingly realistic clips; a 2025 version added sound and a social app, but OpenAI shut the service down in 2026. in 2026. Some researchers see such models as nascent world modelsworld model, n.: An AI's internal simulation of how its world works, which lets it predict what will happen next — especially in response to its own actions — and so plan, imagine or practice without acting for real.; others doubt that they learn physics. Their realism sharpens concerns about deepfakesdeepfake, n.: A fake but realistic video, photo or voice recording made with AI that puts words in a real person's mouth or shows them in a scene that never happened. and AI slopAI slop, n.: Low-quality content churned out in bulk by AI — bland articles, fake-looking images, filler videos — posted to grab clicks or fill space, with little care for accuracy or value..

The text-to-video model kept the cat consistent for eight seconds, then gave it a second tail.