The rapid improvements in AI video generation don't appear to be slowing down any time soon, and Google has announced a major upgrade to its own video maker in the form of Veo 2. It follows the introduction of Veo in May, and OpenAI rolling out its Sora AI video generator to the public earlier this month.
Google says Veo 2 beats Sora and other competitors in terms of user preferences and prompt adherence—you can see the relevant charts for yourself on the Veo 2 website—and that the new model brings with it "enhanced realism and fidelity" as well as better handling for motion and improved camera controls.
Veo 2 has boosted its video-generating capabilities through "an improved understanding of real-world physics and the nuances of human movement and expression," according to Google—so it sounds like you'll see less weirdness in your clips.
The new model also has a better handle on the language of cinema and film. That means prompts can include a specific movie genre, a certain type of lens, or a particular cinematic effect. There's now support for terms such as "low-angle tracking shot," "close-up shot," or "shallow depth of field," which can help generated videos fit more closely to a prompt.
Videos can be created at resolutions up to 4K (an upgrade from 1080p), and can now run to several minutes in length. Some of the example clips Google has been showing off look impressive: We've got drinks being poured, dogs riding inflatables in pools, cartoon characters talking to the camera, and scientists peering down microscopes.
It's not flawless, however, and Google admits Veo 2 still struggles with consistency in complex scenes. This is an issue with AI video generators as a whole—they don't understand the physics and space of the real world, even if they can do a rough approximation of it based on a vast amount of training data. You'll notice that none of these videos show something going out of shot and then coming back in, most probably because the AI would forget about it or render it differently.
The process for getting at Veo 2 is the same as it already was for the first Veo model: You need to head over to Google's VideoFX tool on the web, which is part of Google Labs, and sign up for the waiting list. For now, you need to be in the U.S. to try out Veo 2, and you need to be 18 or older.
Google is promising an expansion of Veo 2 availability over the course of next year—including putting it into YouTube Shorts for creators—but for the time being it's not available in other tools such as Google Gemini (which does offer access to the latest Imagen 3 art generator model).
As I don't currently have access to Veo 2, I can't give you too many details on the limitations that are in place. Google says "each user is limited to a certain number of daily generations" but it's not clear how many videos you can whip up each day, or how long they can be. OpenAI is letting ChatGPT Plus users produce up to 50 videos per month, with the length capped at five seconds and the resolution limited to 720p.
Any videos you produce will come with an "invisible SynthID watermark" that identifies them as having been AI-generated—part of Google's attempts to minimize the risk of Veo 2 being used to deceive viewers and spread misinformation (safety concerns are also part of the reason for the slow rollout, Google says).
