Progress of Video Generation

There have been huge improvements in AI video generation recently, just like there has been for image generation.

One example that has been used as a benchmark for video generation is the Will Smith eating spaghetti test. A benchmark is used to rate how well a model does a task. In this case, video generation models can be benchmarked on how realistic a video of the actor Will Smith eating spaghetti looks.

In 2023, a user of the website Reddit posted videos of Will Smith eating spaghetti generated using a video model tools called ModelScope. The video was shared across social media and this started a trend where model developers tested their new models by generating their own versions of the video.

The video below shows examples from the original post in 2023 and a more recent generated video from 2025 using Google's Veo 3 video model. There are huge differences in how realistic the videos are, even after only two years.

In 2026, Forbes magazine reported that the Seedance 2.0 model had passed the Will Smith eating spaghetti test, meaning it was realistic enough to pass as real video. You can learn more about this test on Wikipedia.

Some reasons for the improvement of video generation include an increase in computing power, more training data and smarter approaches to generation that improve realism.