Text-to-video models are notoriously sensitive to how prompts are phrased. A new preprint on arXiv introduces TIDE (Trajectory-Informed Skill Distillation), a method designed to make prompt optimization more efficient. According to the abstract, iterative prompt optimization typically improves generation quality only at the cost of repeated video generation and vision-language model evaluation.
TIDE's approach is to distill skills from the optimization trajectory, presumably reducing the number of expensive generation-and-evaluation cycles needed. The paper is a single source, and the abstract is cut off before results are described, so the claims here are limited to the problem statement and the proposed method.