Kuaishou's RaG unifies video generation and recommendation, boosting ad revenue by 1.87%

Kuaishou researchers have introduced Recommendation-as-Generation (RaG), a paradigm-shifting framework that unifies personalized video generation and recommendation at industrial scale. RaG moves beyond traditional recommendation systems that match users to a fixed pool of pre-produced videos. Instead, it generates videos on demand based on inferred user interests. The key innovation is Shared Semantic IDs (SIDs), which disentangle video representation into content semantics and creative style semantics. This allows for fine-grained user interest modeling and controllable video generation. The system further uses Video Generation Agents (VGAs) conditioned on inferred SIDs to handle hierarchical planning and refinement—covering visual composition, audio alignment, and artistic effects.

Deployed on a platform with over 400 million daily active users, RaG was evaluated in a revenue-critical advertising scenario. Online A/B tests showed a 1.87% improvement in ad revenue compared to a strong generative retrieval model (GRM) baseline. This demonstrates that combining generative recommendation with video generation can drive meaningful business outcomes. The paper also introduces a synergistic cross-domain reward learning mechanism that jointly optimizes interest alignment, user feedback, and video quality. RaG suggests a closed-loop generative system where user interactions further refine the generation model, creating a flywheel effect for personalization.