Google DeepMind's Veo 2 is here to steal the spotlight from OpenAI's Sora

Last week, OpenAI finally released its long-awaited video generation model Sora in one of the biggest announcements from its ongoing holiday-themed 12-day press engagement. Less than a week later, Google launched Veo 2, a more capable rival model—at least on paper. According to a blog post introducing Veo 2, the model can generate minutes-long video clips in resolutions up to 4K. This means that, theoretically, Veo 2 can produce clips several times (at least six) longer than Sora, in a resolution up to 4x higher, as the latter can produce clips up to 20 seconds long at a 1080p resolution.

The blog post highlights several key Veo 2 features, such as an improved understanding of real-world physics, human movement and expressions, and cinematography concepts. The first two allow Veo 2 to create detailed, realistic-looking videos, while its knowledge of genres, lenses, and cinematographic effects increases controllability. This enables users to ask for details within their prompts such as making the clip look like it was taken using an 18 mm lens, like it has a shallow depth of field.

Finally, Google DeepMind also claims Veo 2 is designed to hallucinate less than its competitors, that is, to avoid producing uncanny-looking scenes or adding unexpected objects like extra fingers. Although there is no evidence of this except for Google's handpicked video clip examples, Veo 2's outputs look impressive. Currently, the model is exclusively available in Video FX, the Google Labs video generation tool. Google plans to expand access to Video FX and eventually bring Veo 2 to power other experiences like YouTube Shorts.