Instead of manually tweaking prompts and settings repeatedly, you can use an AI agent to systematically optimize video generation—first refining what you ask for, then finding the best random parameters—making video synthesis more predictable and reliable.
This paper presents an agentic framework that optimizes image-to-video generation by automatically refining prompts and hyperparameters rather than relying on trial-and-error. It uses a multimodal language model to improve prompts based on semantic checks, then applies Bayesian optimization to find the best random seeds and settings, achieving 69% preference over baseline methods.