Google was busy last week with new AI announcements, including the impressive Gemini 2, Deep Research, and new features in NotebookLM.

Now it's the turn of the search giant's more creative side, as DeepMind unveiled a new version of its Veo video model. First announced at Google I/O earlier this year, Veo is one of the best AI video generators, competing directly with OpenAI's Sora.

Veo2 not only improves visual realism, but also provides a better understanding of physics and ensures that motion is depicted more accurately. This is similar to the updates made by Pika Labs with the new Pika 2.0 model.

According to Google, Veo 2 achieves state-of-the-art results compared to other leading models, especially for human facial expressions.

The model can be tested with VideoFX and a new lab experiment called Whisk, which uses AI to visualize ideas. It will also be available to developers and companies on Google Cloud.

Google claims that Veo 2 can understand real-world physics. This is the Holy Grail for video models of AI, an area where even the best struggle, including OpenAI's Sora.

I have not tried Veo 2 myself, but the videos Google has shared (including one showing bees surrounding a beekeeper) suggest they may have solved this problem.

Veo 2 also understands different camera types. This is something the image model has had for some time and can be used effectively.

According to Google, you can: prompt for “18mm lens” and Veo 2 knows to create wide-angle shots, which this lens excels at,” adding that you can also prompt for ‘shallow depth of field’ to blur the background.

Veo 2 can generate clips up to one minute in length at 4k resolution. It is trained on the “language of cinematography,” which Google claims results in fewer extra fingers and unwanted objects.

Veo 2 has been added to VideoFX, but the service still operates a waiting list. It will also be added to YouTube Shorts in the future to enable AI content creation on the video platform.