Google Veo 2: Cinematic Controls, Realistic Scenes and Motion Limits

Google DeepMind has announced Veo 2, a text-to-video model, and the examples look impressive to me. I am particularly struck by the realistic appearance and the way some clips depict motion. This review is a snapshot of the December 2024 release; comparisons with Sora and other models refer to the versions and selected samples available then.

The first example is a close-up of a DJ with her eyes closed, moving to the music amid neon colors. The prompt describes her face and hair, a shallow depth of field and the camera’s close framing. I find the expression and image quality convincing. Prompts that specify shot type, lens and camera movement seem especially useful here.

A second sample asks for a medium shot of an animated cartoon girl in a 1980s kitchen, again specifying shallow depth of field. Another uses a low tracking angle and an 18 mm lens to show a muscle car drifting through a street. I like the reflections and tire smoke, and the detailed cinematic instructions give a clear idea of the intended shot.

Google displays a human-preference comparison against Meta Movie Gen, Kling 1.5, Minimax and Sora Turbo. In the chart, raters preferred Veo in 58.8% of the Sora Turbo comparisons, preferred Sora in 26.7%, and tied 14.5%. Those numbers describe Google’s particular evaluation and model versions, not a guarantee that Veo wins on every prompt.

A palace scene shows a woman at a desk reflected in a mirror as the camera moves behind her. I find the reflection and rich interior convincing. The prompt also asks for a crimson gown, letter, background figures and dramatic lighting; some are only glimpsed or represented imperfectly. A polished sample is still one selected output, not a complete test of scene consistency.

Complex physical movement remains difficult. I show a Sora gymnast whose body appears disconnected, followed by a Veo 2 ice skater whose legs and torso distort during a spin. A skateboarder also makes implausible movements after leaving the ground. These examples show that attractive frames do not guarantee consistent anatomy or physical motion throughout a clip.

The breakfast example depicts syrup over pancakes and coffee pouring into a clear cup. I like the appearance of the liquids, but the requested close-up of a cut orange does not appear, and some other prompt details are missing. Visual quality and complete prompt adherence are separate questions.

Another sample follows a beekeeper among hives and sunflowers, with a detailed prompt mentioning camera movement, a 35 mm lens and film stock. I find the farmer and bees convincing, although the requested farmhouse and shutters do not appear. The ability to express camera language in a video prompt reminds me of techniques used in image generators.

Google describes improved realism, motion and camera control as Veo 2’s differentiating features. Its [December 2024 announcement](https://blog.google/innovation-and-ai/models-and-research/google-labs/video-image-generation-update-december-2024/) also notes that generated videos can still contain unwanted details. These are company claims and examples; a model’s apparent realism varies by task and output.

In a side-by-side rhinoceros prompt, the Veo output looks somewhat rigid, while the Sora sample begins with convincing movement but later appears to float. My preference for the Veo clip is a judgment about these particular generations, not a controlled comparison across all prompts. Another Veo rhinoceros sample looks to me like nature-documentary footage.

I look through several horse clips. One horse passes through a fence, even though the rest of the scene looks good; another shot hides the full body, and another has slightly unnatural movement. Some other examples appear more convincing. The comparison illustrates why examining the whole clip matters more than judging a single frame.

I compare additional clips from Hailuo and Sora Turbo. In the samples shown, some animals float, morph or move awkwardly. These are selected examples from different models and are not a comprehensive independent evaluation, but they make the remaining motion problems easy to see.

The later examples range from a pelican riding a bicycle to a car going over a waterfall, a market in the sky and animated potato chefs. These are fantastical rather than physically realistic subjects, yet I find their visual presentation compelling. A drifting car also retains a readable Veo license plate in the sample shown.

I also examine fabric falling onto marble, a woman and cat briefly hidden by a passerby, an octopus, shuffling cards and a jumping cat. The examples show different challenges—cloth movement, occlusion, morphing and hands. I find many impressive, while noting that the octopus and cards are not perfect.

The final clips show running giraffes and a tomato being sliced. I am particularly impressed by the tomato’s reflections and apparent motion. These are my visual impressions of selected examples; they do not establish that the model consistently reproduces animal movement or real-world physics.