I’ve been trying to create long form GenAI films and found myself struggling to generate video of a character to wink. I found myself quite amused by the results.

So I built a benchmark around getting video models to wink.

Every model gets the same prompt and the same starting frame.

Prompt: “She winks with her right eye.”

Models tested: Veo3 / Kling 2.1 Master / Runway Gen-3 / SeedDance 1.0 Pro / Moonvalley Marey / Wan 2.2 / Minimax Hailuo-02 / Midjourney

I've ran 9 random seeds for each model.

It looks like Kling 2.1 Master is doing the best of this batch.

Would love to hear other small, quirky things you've struggled video models to do — I’ll try to add them!