COMIC: Agentic Sketch Comedy Generation

Susung Hong; Brian Curless; Ira Kemelmacher-Shlizerman; Steve Seitz

cs.CVMar 11, 2026

COMIC: Agentic Sketch Comedy Generation

Authors: Susung Hong, Brian Curless, Ira Kemelmacher-Shlizerman, Steve Seitz

Organizations: University of Washington

Abstract

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.

Figures & tables

Figure 1: COMIC takes images, voices, and brief descriptions as input, and automatically generates entertaining comedy scripts along with corresponding video and audio. Our framework flexibly builds stories around multiple characters and custom backgrounds. Each generated comedy video is 1–2 minutes long; please watch them at https://susunghong.github.io/COMIC .

Figure 2: Overall agentic flow. COMIC is loosely modeled on a production studio, with agentic counterparts for a writer, critic, editor, and scene director. The writing and rendering loops generate scripts and videos through island-based competition and iteration.

Comparison	Mean Critic	Single Best	Task-Wise Best
Top vs. Middle	0.557	0.554	0.578
Top vs. Bottom	0.654	0.670	0.716

Table 1: Held-out test accuracy for writing critics.

# Samples	Comparison	Mean Critic	Single Best	Task-Wise Best
0	Top vs. Middle	0.542	0.572	0.642
0	Top vs. Bottom	0.700	0.728	0.802
15	Top vs. Middle	0.547	0.590	0.644
15	Top vs. Bottom	0.697	0.702	0.808
45	Top vs. Middle	0.542	0.572	0.644
45	Top vs. Bottom	0.708	0.740	0.830

Table 2: Writing-critic validation accuracy as a function of in-context sample size. Task-wise selection has the highest reported accuracy for each sample size.

Figure 3: Script writing loop. Isolated script populations evolve on separate islands under distinct writing-critic committees sampled from the aligned pool. Losing scripts are refined through round-robin pairwise tournaments.

Figure 4: Video rendering loop. Script-conditioned rendering critics refine scene directions, while shot- and video-level tournaments select from refinement histories and scene realizations, i.e. , depth and breadth, respectively.

Method	Funniness↑	Watch More↑	vs. Human↑	Script↑	Narrative↑	Realism↑	Consistency↑
Veo 3.1 ( Google DeepMind, 2025 )	2.32	2.36	2.27	2.18	3.32	4.91	5.05
Sora 2 ( OpenAI, 2025 )	2.73	2.73	2.32	2.45	3.36	5.73	5.50
VGoT ( Zheng et al., 2024 )	1.18	1.27	1.14	1.00	1.23	2.00	2.32
MovieAgent ( Wu et al., 2025 )	1.27	1.09	1.18	1.09	1.09	1.27	1.14
COMIC (Ours)	3.45	3.09	3.05	3.32	4.50	4.27	4.50

Table 3: Human evaluation of generated videos. Scores are means on 7-point scales. vs. Human compares each average, non-curated output with professional, human-made sketch comedy.

Method	Single Best			Channel-Wise Best
Method	Win Rate	Inter-Diversity	Intra-Diversity	Win Rate	Inter-Diversity	Intra-Diversity
Veo 3.1 ( Google DeepMind, 2025 )	0.010	0.308	0.369	0.105	0.263	0.360
Sora 2 ( OpenAI, 2025 )	0.075	0.531	0.722	0.175	0.310	0.563
VGoT ( Zheng et al., 2024 )	0.000	0.000	0.000	0.010	0.105	0.189
MovieAgent ( Wu et al., 2025 )	0.000	0.000	0.000	0.130	0.088	0.180
COMIC (Ours)	0.440	0.780	0.682	0.390	0.519	0.693

Table 4: Automated evaluation by benchmark video critics independently sampled and separately selected from COMIC’s optimization critics, averaged across channels. Single Best uses one top benchmark critic; Channel-Wise Best aggregates per-channel benchmark critics.

Figure 5: Effect of generation, measured by Win Rate , Inter-Diversity , and Intra-Diversity across iterations relative to the 0th generation. Average values are presented as solid lines, and the range is depicted as shading. Through the 4th generation, the win rate increases substantially. Inter-Diversity (diversity across scripts) initially drops due to the emergence of coherent, favorable responses but increases as generations progress, driven by our divergent mechanism.

Figure 6: Qualitative examples of corrections made by script-conditioned rendering critics, including character, background, framing, and visual-artifact fixes.

Benchmark critic	Metric	VGoT ∗	MovieAgent ∗	No-Refine ∗	COMIC (Full) ∗
Single Best	Win Rate $\uparrow$	0.03	0.02	0.30	0.44
Single Best	Inter-Div. $\uparrow$	0.62	0.46	0.70	0.78
Channel-Wise Best	Win Rate $\uparrow$	0.17	0.13	0.31	0.39
Channel-Wise Best	Inter-Div. $\uparrow$	0.37	0.33	0.50	0.52

Table 5: Controlled-backbone automated evaluation. Starred rows indicate COMIC backbones. No-Refine is an ablation of our framework that removes multi-island iterative refinement.

Figure 7: Single- and multi-island settings.

Figure 8: Scale.

Figure 9: Human preference study against the critic-free ablation.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 10: Additional sketch comedy videos featuring varied generated situations.

	Comparison	Studio C	FAH	VLDL	K&P	SNL	Average
Mean Critic	Top vs. Middle	0.63	0.40	0.58	0.53	0.59	0.55
Mean Critic	Top vs. Bottom	0.72	0.62	0.79	0.60	0.81	0.71
Single Best	Top vs. Middle	0.62	0.52	0.56	0.57	0.59	0.57
Single Best	Top vs. Bottom	0.67	0.70	0.84	0.69	0.80	0.74
Task-Wise Best	Top vs. Middle	0.70	0.52	0.70	0.65	0.65	0.64
Task-Wise Best	Top vs. Bottom	0.85	0.78	0.85	0.72	0.95	0.83

Appendix

Table 6: Channel-specific writing-critic validation accuracy. Task-wise selection has the highest average accuracy across the reported channels and engagement tiers. FAH = Foil Arms & Hog; VLDL = Viva La Dirt League; K&P = Key & Peele.

Figure 11: Logistic growth model fits for cumulative view counts across different comedy channels.

Component	Model
Concept and Script Writing	Claude 4 Sonnet
Language-Only Writing Critics	Claude 4.5 Haiku
Rendering / Benchmark Video Critics	Gemini 3 Flash Preview
Script Inspection	Claude 4.5 Opus
Scene Direction Generation	Claude 4.5 Opus
Meta-Critic	Claude 4.5 Opus

Appendix

Table 7: Models used in the pipeline.

Scale	$K$	$\|S_{k}\|$	$\|\mathcal{C}_{k}\|$	$D$	$\|\mathcal{C}_{\text{render}}\|$
Small	2	2	2	1	0
Base	3	3	3	2	1
Large	4	4	4	4	2

Appendix

Table 8: Scale configurations for the COMIC framework.

Figure 14: Distributions of human evaluation ratings for COMIC compared to baseline methods.

Explore similar work

VideoWeaver: Evaluating and Evolving Skills for Agentic Long Video Generation
Jun 6, 2026Jianhui Wei, Yan Zhang, Jie Tan +7Video Agent Long Video Generation
Zhejiang University · ByteDance
CineForge: Self-Improving Agents for Long-Horizon Video Generation
Aug 30, 2026Junxiang Liu, Lin Wang, Haiyu Shi +10Video Agent Video Generation
Tongji University, Shanghai, China · Kuaishou Technology, Beijing, China · Fudan University, Shanghai, China
Bridging Creative Intent and Visual Quality: Creator-Driven Recurrent Video Generation with Agentic Feedback Loops
Jun 17, 2026Denis Savytski, Aiden Lei, Heding Liu +4Video Generation Generative Artificial Intelligence
Department of Computer Science, University of California, Davis, USA · The Harker School · Basis Independent Silicon Valley +1

Jun 6, 2026, cs.CVJ/K move · Enter open · S save

VideoWeaver: Evaluating and Evolving Skills for Agentic Long Video Generation

Jianhui Wei, Yan Zhang, Jie Tan, Hengchuan Zhu+6

Zhejiang University · ByteDance

Video Agent Long Video Generation Long Videos Skill Evolution

Open paper AskPDF arXiv

Agentic long video generation requires planning, tool orchestration, and cross-clip coordination over a long horizon. Most existing video agents either rely on static, human-crafted workflows, which require substantial manual effort and poorly adapt across tasks, or iteratively refine the output of the current task without persistently distilling execution experience into reusable skills for future tasks. We introduce VideoWeaver, an agent harness and benchmark that evaluates and evolves skills for long video generation. Given a single high-level instruction, an agent dynamically composes foundation skills into its own workflow rather than following a predefined pipeline. We construct a benchmark of 16 task categories and 285 cases, with references spanning text, image, audio, video, and their combinations. We further propose an evidence-grounded agent-as-judge that inspects both the execution trace and the final video to diagnose process and output failures. Based on this feedback, our evolution algorithm progressively refines category-level composition and creator skills, allowing recurring experience to guide dynamically constructed workflows for unseen cases. Experiments show that explicit composition skills improve the generation process over foundation skills alone, while skill evolution further improves output quality and generalizes to unseen cases. Incorporating judge feedback yields additional gains, especially on output metrics, and the agent-as-judge aligns well with human, particularly on process metrics. Code is available at https://github.com/JianhuiWei7/VideoWeaver.

Abstract

Explore similar work

VideoWeaver: Evaluating and Evolving Skills for Agentic Long Video Generation

CineForge: Self-Improving Agents for Long-Horizon Video Generation

Bridging Creative Intent and Visual Quality: Creator-Driven Recurrent Video Generation with Agentic Feedback Loops