Google DeepMind Unveils Veo 2, Its Latest AI Video Model to Rival OpenAI’s Sora
Google DeepMind has introduced Veo 2, its latest AI video generation model, just days after OpenAI’s widely discussed Sora made headlines. The unveiling marks Google’s determined push to take the lead in the rapidly evolving AI video landscape.
Veo 2 boasts the ability to produce videos at up to 4K resolution – a significant improvement over its predecessor, Veo, which capped at 1080p. The model introduces advanced capabilities like enhanced “camera control” and improved physics simulation, allowing for more lifelike motion, detailed object interactions, and cinematic effects, such as sweeping pans and close-ups, all generated through user prompts.
In head-to-head comparisons, Google DeepMind claims Veo 2 holds a competitive edge. During user preference tests, 59% of human evaluators favored Veo 2 over OpenAI’s Sora Turbo – a faster version of Sora. The model also performed strongly against other rivals like Meta’s Movie Gen and Minimax, with its only notable dip occurring against China’s Kling v1.5.
Despite its impressive potential, Veo 2’s current capabilities remain somewhat restricted. While the model can theoretically generate two-minute 4K videos – six times longer and four times sharper than Sora – its output in Google’s experimental VideoFX tool is presently limited to 720p resolution and eight-second clips. OpenAI’s Sora, on the other hand, generates up to 20-second clips at 1080p resolution. Google says it is actively working to scale Veo 2’s output to unlock its full potential.
Veo 2 also grapples with a common AI video challenge: maintaining coherence in complex scenes. For instance, the model struggles with intricate human movements, such as gymnastics sequences – an issue that also affects competitors like OpenAI’s Sora and Runway’s Gen-8 Alpha.
To address concerns of misuse, DeepMind has incorporated its SynthID watermarking technology into Veo 2. This invisible digital watermark helps trace AI-generated content, a step the company hopes will promote responsible usage of the tool.
However, questions remain about Veo 2’s training data. While DeepMind has not disclosed its video sources, speculation points to YouTube – Google’s own platform – as a likely contributor.
The model is currently available on Google Labs’ VideoFX tool for U.S. users on a waitlist. Alongside Veo 2, DeepMind also introduced updates to its Imagen 3 text-to-image model, enhancing image quality and stylistic accuracy for its ImageFX platform, now accessible in over 100 countries.
With its advancements in resolution, cinematic realism, and motion physics, Veo 2 signals Google’s commitment to driving innovation in AI video generation. Although hurdles remain, its scaling potential positions it as a formidable rival to OpenAI’s Sora and other AI video tools.
