Google Deepmind announces a new version of its AI video generator Veo. Veo 2 cranks up video quality and has a better understanding of physics and cinematography.
The first version of Veo, an AI model that can generate short video clips based on text prompts, has only just been launched or Google Deepmind is already announcing its successor. Veo 2 is being made available in a limited test version through the video platform VideoFX. The new model can create much more realistic videos.
The new model has a better understanding of the laws of physics, Google Deepmind researchers explain in a blog. As a result, it can better capture movements and expressions. It also cranks up the image quality and length of videos. Veo 2 can create images at 4K quality and is no longer limited to clips of up to one minute. The test version still sticks to a modest 720p resolution, notes TechCrunch.
Cinematographic knowledge
Veo 2 has also been given a cinematography course to master the “unique language of cinema,” Google writes in a blog. You can ask the model to adopt a style, use a specific type of lens or apply cinematic effects. For example, the model should understand that if you ask it to use an 18 mm lens, it should take a wide-angle shot.
In the blog, Google shows some carefully selected examples to demonstrate the possibilities with Veo 2. At the same time, it is also honest about limitations. While the possibility of hallucinations, such as adding an extra finger, should be significantly reduced, it cannot be ruled out that Veo 2 adds unnatural things.
Competing with Sora
With Veo 2, Google wants to take on OpenAI once again. Google was too quick to outdo OpenAI do it Veo officially launched in early December. A reaction was not long in coming and only a few days later OpenAI put its video generator Sora under the Christmas tree (but not in Europe). Now Google is trying to take the spotlight back to itself with Veo 2.
read also
OpenAI’s video AI Sora can’t handle popularity at launch
AI video seems so quietly poised for a breakthrough. This is not entirely without risk, because the better the imagery of Veo and Sora gets, the harder it becomes to distinguish videos made by humans from AI-generated videos. Google is trying to solve this by adding an invisible watermark to videos created by Sora.
Then again, what Google remains vague about is how it trained Veo. Chances are very plausible that this was done with YouTube’s vast database, which is owned by Google. The Deepmind lab and parent company Google cite the principle of “fair use” for training models with public data.
Google is active not only in video, but also in AI-generated still images. It launched Imagen 3 earlier this month and is now even unveiling a brand new image generator Whisk.