A text-to-video model cannot read characters directly, so the prompt is first broken into tokens. Most systems use subword tokenization: common words become single tokens, while rare or novel words are split into smaller reusable pieces. This lets a fixed vocabulary cover words the model never saw during training, and it keeps related forms such as "drift", "drifting", and "drifted" sharing a common stem token. The result is a short sequence of integer token IDs, which is the actual input to the text encoder.
How Text-to-Video AI Works Under the Hood