acceptodds
Withdrawn from ICLR 2027

COMIC: Agentic Sketch Comedy Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

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.

Primary area:
generative models
Keywords:
Open-Ended Video Generation, Evolutionary Optimization, Multi-Agent Systems

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ADRwOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

COMIC is a fully automated multi-agent system that turns character and background references into 1–2 minute sketch comedy videos. It aligns LLM writing critics with YouTube engagement data, evolves scripts through island-based pairwise tournaments, and renders videos with script-conditioned critics and shot- and video-level tournaments.

Critical strengths
  • An interesting and very open-ended task. Generating genuinely funny long-form video is hard, and the system covers the whole production pipeline from concept and script to storyboard, audio-video synthesis and continuity.

  • Grounding critics in real audience data. Selecting LLM critics by how well they predict real viewer engagement, rather than hand-writing judge prompts, is a reasonable idea.

Critical weaknesses
  • The aligned critics are barely above chance. On held-out data, the selected critics reach only 0.578 on Top vs. Middle, where chance is 0.5, and 0.716 on Top vs. Bottom (Table 1). So they can hardly tell a hit sketch from an average one, yet the whole script evolution relies on them as judges. The selection is also noisy, with only about 10 scripts per tier per channel in each split.

  • The effect of alignment is never shown end to end. The engagement-aligned critic is the key contribution, but there is no ablation that swaps aligned critics for unaligned ones (e.g., zero-shot or random personas) in the full pipeline. The "No Critics" study (Fig. 9) removes critics entirely, which is a different question.

  • The automated evaluation is close to circular. The benchmark video critics use the same model (Gemini 3 Flash), the same meta-critic procedure and the same engagement data as the rendering critics that COMIC optimizes against. Being "independently sampled" does not remove shared biases.

  • The "diversity" metrics do not measure diversity. Both metrics are variances of win probabilities (lines 348–349), and Intra-Diversity is essentially disagreement among evaluators on the same video. No semantic or stylistic diversity of scripts is measured, even though diverse comedic styles are a main selling point of the island design.

  • The human study is too small. Only 22 raters and 4 videos per method, which is not enough to support the conclusions.

Questions for authors
  • What happens if the aligned critics are replaced with unaligned ones (e.g., zero-shot personas) in the full pipeline?

  • Can you evaluate with judges from a different model family, or with a larger human study?

  • Can you add Seedance 2.5 as a frontier baseline?

Reviewer_Xz6MOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper targets fully automated generation of short comedic sketch videos from character and background references, a setting where humor is subjective and open-ended and current video models only produce short clips without cross-shot consistency. It proposes COMIC, a multi-agent pipeline modeled on production-studio roles, with three algorithmic contributions: LLM-based writing critics aligned with real viewer preferences by analyzing a corpus of YouTube comedy videos and their engagement; multi-island relativistic evolution, where script populations on separate islands, each with its own critic committee, improve through round-robin tournaments; and script-conditioned rendering critics running shot- and video-level competitions over rendered candidates. Experiments report higher scores than agentic and video-generation baselines under both human evaluation and independently selected video critics.

Critical strengths
  1. Humor evaluation is grounded in external data rather than hand-written prompts: writing critics are generated and selected against engagement scores from 4,940 YouTube sketch videos, with held-out accuracy reported. This makes a subjective objective auditable.

  2. The multi-island topology preserves distinct comedic niches instead of collapsing to a single objective: each island evolves under its own critic committee, and the critic-guided update merges crossover and mutation in one step.

Critical weaknesses
  1. The proposed framework requires multiple-generation budget and iterative refinement, while Sora 2 and Veo 3.1 are compared at their default single-generation budget. Without a compute-matched baseline, the contribution of the agentic loop versus sampling more is unclear. Besides, the baselines are also dated, newer generators such as Seedance and MiniMax H3 are absent, so whether the loop still adds value on stronger backbones is unknown.

  2. The generated sketches are not yet convincing: win rates stay below 0.5 even against middle-tier human sketches, and the supplementary videos look monotonous in visuals, voice, and camera work and are rarely funny. Since these are presumably the strongest outputs and humor is the stated objective, gains in script or consistency metrics do not compensate.

Questions for authors
  1. How does COMIC compare against a compute-matched baseline with N samples and best-of-N selection at a comparable budget?
Reviewer_Y3zCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

COMIC generates short comedy videos from character and background specifications. It constructs writing critics by selecting critic prompts whose preferences correlate with projected YouTube engagement. Script populations evolve on separate islands through pairwise comparisons and feedback-driven rewriting.

Selected scripts are converted into shots using a production pipeline with character assets, continuity memory, and script-conditioned rendering critics. Shot-history and scene-level tournaments select final outputs. The paper reports automated and human comparisons against agentic video systems and frontier video generators.

Critical strengths

COMIC’s strongest contribution is its attempt to make automated evaluation useful for an inherently subjective creative task. Rather than relying entirely on manually assigned critic personas, the authors generate and select writing critics according to their agreement with audience-engagement signals from existing comedy videos. Although engagement is an imperfect proxy for humor, this gives critic selection an external basis and offers an interesting direction for other creative domains. The use of separate script populations governed by different critic committees is also a sensible way to accommodate competing comedic preferences without forcing every candidate toward the same style. The system carries this iterative approach into production through script-conditioned rendering critics and selection from earlier shot revisions, allowing it to recover from refinements that make an output worse. Table 5 provides some supporting evidence: with the generation backbones controlled, the full system improves the automated win rate from 0.30 for the no-refinement variant to 0.44 under the Single Best benchmark critic.

The novelty is primarily in combining audience-calibrated critics, preference-diverse script search, and audiovisual refinement into a complete workflow. Island-based evolution and feedback-driven rewriting are established ideas, but their integration here raises worthwhile research questions about how to search and evaluate creative outputs when no single objective adequately captures quality.

Critical weaknesses
  1. Audio and duration differences are particularly consequential for comedy. A system-level advantage under unequal production conditions does not isolate the proposed search method.

  2. Views depend on exposure, topic, channel history, celebrity, and other factors. The logistic correction does not establish that these confounds are removed. Writing-critic accuracy of 57.8% on top-versus-middle examples suggests limited discrimination between closer-quality scripts.

  3. Variance in comparative win probabilities may reflect uneven quality or evaluator disagreement. It does not establish distinct premises, comedic structures, or styles.

  4. Human evidence has limited output coverage and uncertain statistical structure. There are four videos per method and 110 reported responses across five methods. Independent rater counts, repeated-measures structure, and uncertainty require clearer reporting. Absolute scores also remain modest: funniness is 3.45/7 and willingness to watch more is 3.09/7.

Questions for authors

How many independent participants rated each video, how were responses aggregated, and what uncertainty remains when both videos and raters are treated as sampling units?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.