Google DeepMind unveils Veo, a text-to-video generation model

At Google I/O 2024, Google DeepMind unveiled Veo, a model that generates video from text and image prompts. The company said it could produce 1080p footage “in a wide range of cinematic and visual styles” running beyond a minute, positioning it as its answer to OpenAI’s Sora, which had been announced in February but was still unreleased.

Google described Veo as building on its earlier video work — Lumiere, VideoPoet and Imagen Video — and emphasised the model’s grasp of cinematic language and physical consistency: it said Veo could interpret terms like “timelapse” or “aerial shot” correctly and keep people, animals and objects moving coherently across a shot rather than warping between frames, a common failure mode in earlier video generators. As with OpenAI’s Sora demonstrations, the clips Google showed were curated examples rather than results from independent testing, and neither the training data nor a technical report accompanied the announcement.

Consistent with the caution surrounding Sora, Google did not ship Veo as a public product. Access launched through a waitlist for VideoFX, a limited private-preview tool, with Google saying capabilities would extend later to YouTube Shorts and other products. Every video Veo generated carried Google’s SynthID digital watermark, a provenance measure the company had already applied to Imagen-generated images, extended here to a modality where synthetic content would be harder for viewers to distinguish from real footage.

Veo arrived the same day as Google’s Project Astra prototype, part of an I/O keynote that leaned on prototypes and previews across video, voice-and-vision assistance and search rather than immediately shippable products — a pattern that drew comment given how far behind some of it lagged rivals’ public releases.