VEO3 "Slice Some Glass Fruit"

Ask an AI video generator to render something that does not exist and you find out fast how well it understands the physical world. That is the idea behind this quick VEO3 test: type in "slice some glass fruit" and see what comes back.

The prompt and the payoff

The setup is deliberately minimal. Google's VEO3 model gets a four-word instruction and nothing else, no camera notes, no lighting cues, no material specification beyond the word "glass." What it returns is a gloved hand cutting into a red glass fruit that behaves, visually, like something between glass and flesh. It is a useful stress test because it asks the model to reconcile two things that do not usually appear together: the optical properties of glass and the cutting motion of slicing produce.

Why transparent materials are hard for AI video

Transparent and reflective materials are notoriously difficult for generative video models. Light has to pass through the object, bend, and pick up color and reflections from around it, all while the object is moving or being deformed. A cut fruit also needs to look structurally convincing as it separates. Watching how VEO3 splits the difference between believable glass and believable fruit is where the interest of a twenty-second clip like this lives.

Key takeaways

• VEO3 was prompted with a short, direct instruction: "Slice some glass fruit."
• The output shows a gloved hand cutting into a red glass fruit.
• Transparent materials and cutting physics are a genuine test of how an AI video model simulates light and structure.
• A simple, four-word prompt can reveal a model's strengths and blind spots without any elaborate prompt engineering.

Try it yourself

If you want to see how a text-to-video model handles material properties, start with a short, plain-language prompt like this one and watch how it renders transparency and motion together. Humanitarians AI runs quick generative tests like this as part of its exploration of AI creative tools.