Google recently launchedVideo generation model Veo 2 Not sure if you still remember, but Veo 2 is currently only available to a select group of people, so you might not have a clear idea of what Veo 2 actually produces. However, if you’ve been following video generation models, you’ve probably seen users on X who already have access to Veo 2 sharing the videos they’ve generated, which gives those without access a chance to get a sneak peek. Recently, an Emmy-winning Korean motion graphic designer shared their experience of using Veo 2 firsthand to create videos, summarizing the areas where Veo 2 currently excels as well as the drawbacks it still has. If you’re interested in learning more, keep reading!
An Emmy Award winner conducted an in-depth review of Google’s Veo 2 video generation AI and concluded it outperforms all competitors.
A Korean motion graphics designer who recently won an Emmy Award 김그륜 Gryun KimGryun Kim (hereinafter referred to as Kim) posted on YouTube his review after using Veo 2. Kim said he has tested Veo2’s performance and limitations countless times, producing hundreds of videos. After testing, Kim believes Veo 2 is the best-performing AI video generation framework available so far, and Veo 2 performs quite well in handling various scenes and realistic effects. 
First, there’s Veo 2’s performance in simulated scenarios—such as the way animal fur looks dripping wet after getting soaked, the ripples in the water, the bubbles, the cat’s movements, and the buoyancy of the ball in the water. All of these left Kim, a CG expert, stunned.
Kim said the best way to test simulated scenarios is to have AI generate videos of a hydraulic press crushing objects made of different materials. The experiment used Veo 2 to generate videos of a hydraulic press crushing iron kettles, marble, glass, and other materials, and found that Veo 2 handled the shattering and bending effects of objects being crushed very well.
Apart from its basic performance, Veo 2 also handles simulation scenarios with mixed materials quite well. For example, when generating a simulation of a water-filled rubber toy duck being squashed, Veo 2 can realistically depict the rubber compressing while also capturing the water being squeezed out.
Also, with the hydraulic-pressed orange and chocolate, Veo 2 can show the orange juice changing color after it gets on the chocolate once the orange is crushed. If these videos were paired with vivid music, most people would probably have a hard time telling they were AI-generated.
Unrealistic scenario
Veo 2 can handle not only scenarios that might occur in everyday reality, but also ones that don’t exist in the real world. For example, it can generate a penguin wearing blue swimming goggles and a small yellow backpack filled with colorful balloons, walking toward the camera in a bustling city square. The penguin then leaps into a shallow fountain, sending water splashing in all directions.
It is also possible to use Veo 2 for commercial or advertising purposes. Kim believes that based on Veo 2’s current performance, the videos it generates are better and more cinematic than those from any other AI video generation model. For example, by entering the desired background, camera movement path, and logo material in the prompt, Veo 2 can produce stunning results that look like CG.
In addition to using Veo 2 to create advertising logos, Kim also used Veo 2 to produce a Lego Christmas story stop-motion animation. Kim said that based on the results Veo 2 delivers, in the future it might only take one to two days to complete a Lego short animation.
Veo 2 and Sora’s performance under the same prompt.
People might be curious about how Veo2 and Sora compare when given the same prompt. Kim said that for fairness, they used prompts that performed well on Sora and fed them into Veo2 to see its performance. The first scene was a cat skateboarding. Although the video generated by Sora was also realistic, Veo2 was still better in terms of lighting, motion, and realistic physics.
The second scene is the seaside lighthouse scene, and Sora’s rendering is quite good, but if you look closely, you’ll notice that Sora doesn’t fully grasp real-world physics. For instance, videos generated by Sora sometimes appear choppy, and although there is a lot of foam around the lighthouse, the waves are very small, which doesn’t match what you’d see in real life. In comparison, Veo2 generates waves and foam around the lighthouse that are more consistent in scale. 
Kim also found that Veo 2 handles lighting in videos more naturally. With the same prompt, videos generated by Sora felt like they were shot in a studio with set-up lighting, while Veo 2’s results were more natural.
However, Sora isn’t completely without merits. For example, Sora’s Remix feature can help adjust the intensity of character deformation in videos and also improve video quality. Sora also allows for more adjustments to video content through options like looping, editing, and storyboards. Compared to the current Veo 2, Sora’s features are indeed superior.
Veo 2’s disadvantages
Kim mentioned several drawbacks of Veo 2, such as the fact that Veo 2 currently cannot directly use external images for image-to-video generation. You have to use Veo 2’s Text to Image to Video feature, first generating an image with Veo 2, then using that image to create a video. This results in a more uniform style in Veo 2’s output, making it unable to generate a broader range of styles.
Additionally, if Google wants to make Veo 2 usable for commercial purposes or real-world work, it will need to fix the issue of inconsistent characters that Veo 2 generates. Based on Kim’s testing, Veo 2 still produces AI-specific distortion in fast-action segments, with characters almost always coming out somewhat deformed, making it very difficult to create consistent characters. If they want to use Veo 2 for short films, music videos, commercials, and similar productions, the character inconsistency problem definitely needs to be resolved.
Although Veo 2 still has some issues, such as motion errors and character inconsistencies, Kim believes the quality of the videos Veo 2 generates is still unique. Given the pace of AI development, Veo 2 will likely resolve these issues soon. Besides analyzing the performance of Veo 2 and Sora, Kim also teaches viewers how to input more detailed prompts that are better suited to video generation models. Those interested can watch the video below to learn more:
Friends interested in learning about Google’s new video generation model Veo 2 can also click the link below to read more related reports:
Source: KOCPC Chinese