Google’s new AI video model sucks less at physics
Google may have only recently begun rolling out its Veo generative AI to enterprise customers, but the company is not wasting any time getting a new version of the video tool out to early testers. On Monday, Google announced a preview of Veo 2. According to the company, Veo 2 “understands the language of cinematography.” In practice, that means you can reference a specific genre of film, cinematic effect or lens when prompting the model.
Additionally, Google says the new model has a better understanding of real-world physics and human movement. Correctly modeling humans in motion is something all generative models struggle to do. So the company’s claim that Veo 2 is better when it comes to both of those trouble points is notable. Of course, the samples the company provided aren’t enough to know for sure; the true test of Veo 2’s capabilities will come when someone prompts it to generate a video of a gymnast’s routine. Oh, and speaking of things video models struggle with, Google says Veo will produce artifacts like extra fingers “less frequently.”
Separately, Google is rolling out improvements to Imagen 3. Of its text-to-image model, the company says the latest version generates brighter and better-composed images. Additionally, it can render more diverse art styles with greater accuracy. At the same time, it’s also better at following prompts more faithfully. Prompt adherence was an issue I highlighted when the company made Imagen 3 available to Google Cloud customers earlier this month, so if nothing else, Google is aware of the areas where its AI models need work.
Veo 2 will gradually roll out to Google Labs users in the US. For now, Google will limit testers to generating up to eight seconds of footage at 720p. For context, Sora can generate up to 20 seconds of 1080p footage, though doing so requires a $200 per month ChatGPT Pro subscription. As for the latest enhancements to Imagen 3, those are available to Google Labs users in more than 100 countries through ImageFX.