Prompt adherence compares the requested conditions with the produced output. For video, this includes subjects, counts, relationships, actions, camera movement, lighting, dialogue, timing, text, and negative constraints.
A result can look attractive while ignoring essential instructions. Evaluation should separate visual quality from adherence and test the exact shot requirements, since a strong anonymous leaderboard score may not predict success on a creator's specific prompt.
ELI5
Prompt adherence means the AI result follows what the prompt actually asked for. A beautiful output is not successful if it leaves out the important subject, action, or constraint.
For example, a prompt may request two red cars passing under a bridge at night. A cinematic clip with one blue car may look good but has poor prompt adherence because it misses several required details.
How can prompt adherence be measured?
Reviewers or automated checks can score whether each requested subject, attribute, relation, action, style, timing, and constraint appears correctly.
Can a visually impressive generation have poor prompt adherence?
Yes. A result can be attractive while omitting, changing, or contradicting important requirements.
