Text-to-video generation produces a moving video clip from a text prompt, a still image, or both. It is the natural extension of text-to-image generation, but a much harder one. A still image only has to look right once; a video has to stay coherent across many frames, so objects must persist, motion must be physically plausible, and the camera and lighting must behave consistently. Small per-frame errors that would be invisible in a single picture become distracting flicker and warping over time.
The modern wave began with diffusion-based systems. Runway, a co-creator of the latent diffusion architecture, shipped Gen-1 in 2023 for video-to-video restyling and then Gen-2, which it describes as a multimodal system that can generate novel videos from text, images, or clips - including pure text-to-video “using nothing but words.” A central idea, echoed across the field, is that a model trained to predict the next frame of video can acquire a deep, implicit understanding of the visual world, much as next-token prediction did for language. OpenAI’s Sora (2024) and Google’s Veo pushed clip length, resolution, and physical realism further, with later versions adding synchronized audio.
The technology raced from short, dreamlike clips to minute-long, near-photoreal footage in roughly two years. That speed has made provenance and watermarking - such as Google’s SynthID tags - a live concern, because convincing fake video carries obvious risks for misinformation, fraud, and likeness misuse.
Why business readers should care: text-to-video collapses parts of the cost of producing motion content - storyboards, b-roll, product demos, ads. It also raises hard questions about consent, copyright, and trust in moving images, which is why the same labs shipping the models are also shipping detection and labeling tools.