Researchers Introduce AgenticGen for AI Ad Video Generation

Researchers have introduced AgenticGen in an arXiv preprint, presenting a reward-guided agentic framework designed specifically for creating effective advertising videos aligned with online business metrics. The system decomposes generation into two trainable reasoning stages: strategy selection and draft generation.

The framework uses reward models trained on accumulated business feedback and human quality standards, optimizing agent policies through Direct Preference Optimization (DPO) and Group Relative Policy Optimization (GRPO).

Online A/B testing within TikTok's advertising environment showed that AgenticGen outperformed the standard SFT baseline, lifting the click-through rate (CTR) by 2.72%, conversion rate (CVR) by 2.63%, and Advv by 9.61%.