COMIC: Agentic Sketch Comedy Generation
Researchers' new AI system generates SNL-style sketches using competing agents and LLM critics trained on YouTube humor.
A research team from the University of Washington and Google has introduced COMIC, a fully automated AI system designed to generate short comedic videos akin to Saturday Night Live sketches. The system begins with character references and employs a multi-agent framework where different AI agents assume roles loosely based on real production studio positions like writers, directors, and editors. These agents engage in iterative cycles of competition, evaluation, and improvement to optimize both the quality and diversity of the generated content.
A central technical contribution is the development of LLM (Large Language Model) critics specifically aligned with real viewer preferences. These critics were trained by analyzing a large corpus of comedy videos from YouTube, enabling the system to automatically assess and score humor—a notoriously subjective and difficult task for AI. According to the paper, experiments demonstrate that COMIC's output approaches the quality of professionally produced sketches while achieving state-of-the-art performance in automated video generation, marking a significant step toward AI systems that can understand and replicate complex creative processes.
- Uses a multi-agent framework with specialized roles (e.g., writer, director) that compete and iterate to generate ideas.
- Features LLM critics trained on YouTube comedy data to automatically evaluate humor, aligning AI with human preferences.
- Produces short videos approaching professional sketch quality, advancing the state-of-the-art in automated video generation.
Why It Matters
This demonstrates AI's potential to automate complex, subjective creative tasks, moving beyond simple generation to structured, multi-stage production.