COMIC: Agentic Sketch Comedy Generation

The gist: COMIC is a fully automated agentic system that produces short comedic videos similar to sketch shows by employing multiple agents and iterative competition to generate funny scripts and high-quality, consistent video.

How it works

1. The system is loosely modeled on human production studios, using agentic versions of roles such as scriptwriters, editors, and directors (Fig. 2) <ref:2603.11048#pg3>.

2. The overall agentic flow involves a forward pipeline where a writer generates concepts and expands them into full dialogues, a critic evaluates and compares scripts, and an editor revises scripts based on critic feedback <ref:2603.11048#pg6>.

3. Script generation synthesizes a script s∗ ∈ S that establishes a compelling comedic premise, develops it through character interaction, and delivers a satisfying payoff <ref:2603.11048#pg6>.

4. Visual realization translates s∗ into a shot sequence V = [v1,..., vN] that faithfully embodies the narrative while maintaining character identity and scene continuity <ref:2603.11048#pg6>.

Content Optimization via Multi-Agent Iterative Competition (COMIC)

The system addresses the problem of automated sketch comedy video generation by decomposing it into two coupled subproblems: script generation and visual realization <ref:2603.11048#pg6>. The approach is based on the observation that LLMs do occasionally produce humorous content when provided with the right structure, mirroring how real sketch comedy shows operate through brainstorming and iteration <ref:2603.11048#pg6>.

Why Fixed Objectives Fall Short

Traditional goal-based optimization fails because humor is inherently context-dependent and subjective, meaning a fixed scalar objective invites Goodhart’s Law [8] <ref:2603.11048#pg6>. Humor takes many different forms, and it is impossible to aggregate preference profiles into a single, consistent ranking without sacrificing desirable properties <ref:2603.11048#pg6>. Existing agentic strategies are limited because each role is defined by a fixed instruction, meaning the agent always applies the same evaluative lens with no mechanism to explore alternative perspectives <ref:2603.11048#pg6>.

Alignment to Real Viewers

Evaluation quality depends critically on critic selection, which is achieved through a generate-and-select strategy rather than fine-tuning a dedicated critic model <ref:2603.11048#pg8>. This involves synthesizing a large, diverse pool of candidate critics and retaining those whose preferences best align with empirical audience engagement signals <ref:2603.11048#pg8>. Engagement scoring is derived by fitting a per-channel logistic growth model to normalized view counts to project the carrying capacity Lproj, which defines each video’s engagement score <ref:2603.11048#pg8>. Task-specific selection involves defining two comparison tasks, Top vs. Bottom and Top vs. Middle, and selecting the highest-accuracy critic on those pairwise comparisons for each channel and sensitivity level <ref:2603.11048#pg8>.

Script Writing Loop

COMIC utilizes an approach to iteratively evolve a population of scripts by partitioning the global script population into K isolated islands, each governed by a specialized critic committee Ck drawn from Ctask <ref:2603.11048#pg9>. The fitness landscape on island k is shaped by two coupled elements: (1) the island-specific critic committee Ck, which defines evaluative standards, and (2) the evolving script population Sk, which determines the comparative baseline <ref:2603.11048#pg9>. Evolution proceeds through round-robin pairwise evaluation where losing scripts are refined using feedback from winners via an update operator U that integrates comparative feedback with semantic mutation <ref:2603.11048#pg10>.

Video Rendering Loop

The rendering stage translates critic-selected scripts into videos using a competition-based framework for video generation <ref:2603.11048#pg10>. This involves several steps:

1. Script-conditioned critic generation, where a video metacritic agent prender generates critics with diverse personas conditioned on the script <ref:2603.11048#pg10>.

2. Storyboarding, where a scene director agent generates D scene directions that specify rendering of Nj shots in text form <ref:2603.11048#pg10>.

3. Iterative shot refinement with history tournament, where each shot is refined over Crender iterations based on feedback from script-conditioned critics, and a single-elimination tournament selects the best revision across history <ref:2603.11048#pg11>.

4. Test-time scaling via scene-level tournament, which performs a final tournament among the full videos to select the best realization across diverse scene directions <ref:2603.11048#pg11>.

Comparison Studio

COMIC demonstrates superior performance compared to existing baselines when evaluated by human judges across multiple criteria, including Funniness, Watch More, Script Quality, Narrative Quality, Visual Realism, and Consistency <ref:2603.11048#pg24>. COMIC consistently outperforms agentic baselines by large margins across all dimensions <ref:2603.11048#pg24>. Furthermore, the automated ranking (COMIC > Sora > Veo > MA ≈ VGoT) aligns with human results, validating the benchmark as a proxy for human judgment <ref:2603.11048#pg24>. The framework also achieves the highest overall inter- and intra-diversity, demonstrating that its mechanism sustains a diverse range of comedic styles that single-pass methods do not <ref:2603.11048#pg24>.

Scale

The system exposes several scaling dimensions: number of islands K, scripts per island Sk, critics per island Ck, scene directions D, and rendering critics Crender <ref:2603.11048#pg22>. The computational complexity for the writing stage is O(G · K · S k squared · C k) evaluations per island <ref:2603.11048#pg22>. The dominant term for the rendering stage is O(D · N · Crender 2) render calls per script <ref:2603.11048#pg22>. The base configuration runs in approximately one day on a single H200 GPU with an API budget of around 5 dollars, which is orders of magnitude below the production cost of professional sketch comedy <ref:2603.11048#pg22>.

Conclusion

COMIC establishes a new state of the art for fully automated video production by leveraging a multi-island topology where diverse, human-aligned critic committees drive iterative refinement <ref:2603.11048#pg6>. The work connects to the Red Queen hypothesis in evolutionary biology, wherein species must continuously evolve to maintain their fitness against co-evolving competitors <ref:2603.11048#pg6>. Ultimately, this work establishes a new state of the art for automated, engaging, long-form video production.

References

1. Arrow, K.J.: A difficulty in the concept of social welfare <ref:2603.11048#pg16>.

2. Cantú-Paz, E., et al.: A survey of parallel genetic algorithms <ref:2603.11048#pg16>.

3. Chan, C.M., Chen, W., Su, Y., Yu, J., Xue, W., Zhang, S., Fu, J., Liu, Z.: Chateval: Towards better llm-based evaluators through multi-agent debate <ref:2603.11048#pg16>.

4. Cho, H., Ahn, D., Hong, S., Kim, J.E., Kim, S., Jin, K.H.: Tag: Tangential amplifying guidance for hallucination-resistant diffusion sampling <ref:2603.11048#pg16>.

5. Dalal, K., Koceja, D., Xu, J., Zhao, Y., Han, S., Cheung, K.C., Kautz, J., Choi, Y., Sun, Y.: One-minute video generation with test-time training <ref:2603.11048#pg16>.

6. Du, Y., Li, S., Torralba, A., Tenenbaum, J.B., Mordatch, I.: Improving factuality and reasoning in language models through multiagent debate <ref:2603.11048#pg16>.

7. Fernando, C., Banarse, D., Michalewski, H., Osindero, S., Rocktäschel, T.: Promptbreeder: Self-referential self-improvement via prompt evolution <ref:2603.11048!

1. Improve script generation by implementing Island k evolution to explore diverse comedic styles, as this yield[s] a Pareto frontier of diverse comedic styles. This allows for generating sketches across different tonal ranges (slapstick, dry wit, surrealism) rather than converging on a single style.

2. Enhance humor evaluation accuracy by using Task-wise selection of critics based on pairwise comparison tasks like Top vs. Bottom, as this approach was shown to be superior to pooled or single-best critics across all channels and engagement tiers.

3. Increase video rendering quality and consistency by implementing an Iterative shot refinement with history tournament, which uses a SingleElimination tournament across the full history of shots to ensure quality increases after each refinement step, guarding against over-refinement.

4. Scale inference time efficiency by leveraging the multi-island topology for script writing, as this structure makes iterative refinement tractable compared to a globally pooled tournament, and allows for test-time scaling via scene-level tournament.

5. Improve visual coherence and character fidelity by using specialized models like FLUX.2 [dev] alongside TAG [4] for generating canonical character appearances, ensuring that generated video maintains high Visual Consistency.

We propose a fully automated AI system that produces short comedic videos similar to sketch shows such as Saturday Night Live. Starting from character references, the system employs a population of agents loosely modeled on roles in real production studios and structured to optimize the quality and diversity of ideas and outputs through iterative competition, evaluation, and refinement. A key contribution is the introduction of LLM-based critics aligned with real viewer preferences through the analysis of a corpus of comedy videos on YouTube, enabling automatic evaluation of humor. We further propose multi-island relativistic evolution for script improvement, along with script-conditioned rendering critics that perform shot- and video-level tournaments. In both human and automated evaluations, COMIC receives higher scores than agentic and video generation baselines.