Google DeepMind introduced the AI ​​video generator Veo 2, which creates two-minute videos in 4K

Google’s DeepMind division has unveiled Veo 2, a next-generation AI-powered video generator. It creates videos in 4K resolution (4096 × 2160 pixels) up to two minutes long. Thus, it is 4 times faster than OpenAI Sora in resolution and 6 times faster in video length.

Image source: Google DeepMind

At the moment, this advantage, however, is only theoretical: you can try Veo 2 only on the VideoFX experimental site, where the resolution is limited to 720p and the video length is 8 seconds. By comparison, the public version of OpenAI’s Sora generator offers 1080p and 20 seconds. To get started with VideoFX, you’ll have to sign up for a waitlist, although Google has promised to expand its audience this week. In the future, Veo 2 will appear on the Vertex AI platform for business users, but Google did not indicate the exact timing.

Veo 2, like its predecessor, generates videos based on text prompts that can be accompanied by images. Compared to the first generation Veo, the new model “understands” physics better, the image has increased clarity, and the mechanisms for moving the virtual camera have been improved. Veo 2 simulates movement more realistically (for example, depicting coffee being poured into a mug) and the properties of light (shadow and reflection); Different lenses on a virtual camera and cinematic effects are realistically simulated.

The developer also claims that the new model is less likely to exhibit hallucinations: extra fingers or “unexpected objects”; However, as practice has shown, it was not possible to get rid of the “uncanny valley” effect. And in the video of a moving car, upon closer inspection, the road appears to be extremely smooth, pedestrians merge with each other, and some house facades have an appearance that defies the laws of physics.

The new video generator was trained on a large number of videos; DeepMind did not specify which sources were used for this, but one of them is likely the Google-owned YouTube platform. To reduce the risk of deepfakes, Veo 2 has a built-in SynthID system – invisible marking of model-generated videos.

DeepMind also announced that it has improved the performance of the static image generator Imagen 3 – the images created with its help have become brighter, more detailed, and the model itself now more accurately follows the user’s requests. The ImageFX user interface, where the image generator is available, now has drop-down lists that appear directly in the query field and help you achieve more accurate results.