Generating video from a text description has become impressively capable, but it remains a hit-or-miss process. Because each generation is computationally expensive, users often spend significant time and compute iterating until the output matches what they had in mind. A new arXiv paper, ProactGen, tackles this problem by learning to ask clarifying questions before committing to a generation.
Rather than relying on a single prompt to capture all intent, the proposed system proactively seeks additional information. The abstract frames this as a way to reduce the number of expensive generations needed to achieve a satisfactory result. While the full method is not detailed in the abstract, the core idea is clear: clarify first, generate second.
ProactGen is presented as a learning-based approach, meaning the clarification behavior itself is trained rather than hand-coded. This positions it within a broader trend of making generative models more interactive and cost-aware.