AI Video Generation is Fundamentally More Expensive Than Text

The article discusses how AI video generation is more computationally expensive than text generation, due to the inherent complexity of modeling a continuous visual world compared to predicting discrete tokens.

Details

The article argues that the high cost of AI video generation compared to text is not just an optimization issue, but a more fundamental challenge. Text models work well because they can compress meaning into discrete tokens, but video lacks a similar abstraction. Video generation models have to deal with high-dimensional data across many frames, while also maintaining object and motion consistency over time. This makes the problem significantly heavier, as the model has to generate something that behaves like a continuous world and track/maintain a much larger amount of information. This results in higher compute per sample, longer inference paths, and stricter consistency requirements, which quickly add up in cost. Even with model improvements, the underlying structure of the video generation problem may not change easily. The article suggests that meaningful cost reductions will likely require a fundamentally different approach to representing video, rather than just incremental improvements to existing methods.