Google has released its impressive AI video generation model, Veo 2 on Gemini. Not only that, you can generate videos for free using Veo 2 on AI Studio. With this release, I took the opportunity to do a hands-on testing of Veo 2 and see how capable it is at generating realistic AI videos. In this article, I have tested Veo 2’s physics accuracy, motion consistency, spatial and temporal coherence, human realism, and image-to-video capability. And though Veo 2 generates stunning AI videos, not every video is perfect.
Veo 2: Testing Physics Accuracy
For any video generation model, a basic understanding of physics is essential to produce visually consistent videos. So I started my test with a prompt that evaluates Veo 2’s physical understanding. I prompted Veo 2 to create a video of a cat pushing a glass of water and it falling to see how it handles collision impact, gravity, acceleration, and liquid dynamics.
As you can notice, the glass is already tilted even before it’s pushed, and it stays in that state for a long time before it’s knocked, which is impossible. It seems the glass is suspended in the air for a moment, as if there is no effect of gravity or acceleration.
Having said that, the rest of the video is more believable. The liquid spills from the glass realistically after the cat pushes the glass. Overall, I would say Google Veo 2’s understanding of physics is not there yet. While it’s much better than OpenAI’s Sora, which has a poor grasp of physics, Google’s latest video generation model also requires a lot of work.
Veo 2: Testing Motion
Coming to motion, which is another complex concept for video generation models to get right. You see, generating a coherent video of a moving object requires a great understanding of the physical world. And in this test, Google’s Veo 2 video generation model delivers a great result.
I asked Veo 2 to generate a video of a man walking through a forest. In the video linked below, you can see that the character is consistent and visually believable across all frames, which is a huge win for this AI video-gen model.
Not only that, the movement is quite uniform, and the background is almost consistent. Even the falling leaves and squirrels in the background are accurately rendered.
That said, during my testing, I noticed that Veo 2 struggles when there are multiple humans or objects in the scene. It just can’t maintain visual coherency, likely due to limited attention. In any case, for motion and character consistency, Veo 2 wins big time.
Veo 2: Testing Spatial and Temporal Coherence
Spatial and Temporal coherence are critical factors for ensuring that AI-generated videos appear consistent and realistic. Spatial coherence ensures consistency in individual frames, like shadows and reflections. Temporal coherence ensures consistency between frames as the video progresses over time, like maintaining consistent motion and identity.
In my Sora review, I noted that OpenAI’s model struggles more with temporal coherence, resulting in wonky videos. Now, to test Google’s Veo 2 model, I asked it to generate a video of a ball bouncing across a table and colliding with three dice.
Veo 2 generated a fairly plausible video as the ball went past the blue mug on the table, handling the ball’s consistency, shadows, and lighting pretty well. The ball’s motion was also uniform, however, Veo 2 regenerated a completely new frame, losing coherence and relationship among all the objects. The dice moved around unnaturally, and a hand appeared out of nowhere, which broke the temporal coherence.
Nevertheless, Google Veo 2 demonstrates noticeable improvements over earlier video generation AI models. I think Google can fix these issues with iterative updates in the future.
Veo 2: Testing Cinematic Realism
Next, I asked Veo 2 to generate a cinematically realistic video of a man standing in a rainy city. The purpose of this test was to evaluate whether Veo 2 can render believable human faces and nuanced emotions. I also wanted to see how rain is simulated, interacting with the overall environment around it.
And Veo 2 knocked it out of the park. You can check out the video, which is believable and natural. There is a depth of field and focus on the human face. And the rain simulation is absolutely on point.
In the next test, Veo 2 generated a futuristic video of a spacecraft entering the Earth’s orbit. The motion is smooth, and sci-fi elements are visually striking. Veo 2 is very good at producing cinematic videos, and content creators can definitely use this model to embed short clips in their projects.
Veo 2: Testing Image-to-Video Capability
Since Veo 2 supports image-to-video generation, I uploaded a still image of a forest with rivers and mountains. I asked Veo 2 to create a video using this image and add the lush flora and fauna. I also prompted to add a herd of elephants and birds around the riverbed. As you can see in the video below, Veo 2 did a splendid job.
Of course, the elephants magically appear, which is a bit jarring, but other than that, the rest of the video is quite believable. The shadows and motion are executed well, and Veo 2 didn’t alter the base image. I uploaded the same image on OpenAI’s Sora, and it completely failed to create a coherent video in my earlier testing.
Conclusion: Google Veo 2 is Quite Capable But Fumbles in Physics
To conclude, Google’s Veo 2 is the state-of-the-art video generation model and outperforms all competing AI models, including OpenAI’s Sora. At times, it does falter in following the laws of physics, but overall, you often get believable results. Veo 2 is quite capable of generating authentic human faces and emotions.
That said, I noticed that Veo 2 struggles when there are too many human subjects or objects in the scene. In addition, it has a hard time following verbose instructions. Apart from that, Google must be commended for training a powerful AI model that sets a new benchmark in video generation.