ComicScene154: Scene Segmentation in Comics

• ComicScene154 is a manually annotated dataset capturing scene-level narrative arcs in Golden Age comics for multimodal narrative analysis.
• It comprises approximately 600 panels from 34 stories across four genres, providing reproducible benchmarks for scene segmentation.
• The dataset leverages a two-stage, zero-shot LLM pipeline and inter-annotator agreement metrics to validate narrative boundary detection.

ComicScene154 is the first publicly released, manually annotated dataset capturing scene-level narrative arcs in Western comics. Developed to address the absence of fine-grained, scene-based segmentation in comics datasets, ComicScene154 is positioned to advance computational analysis in multimodal narrative understanding. It encompasses annotated data derived from public-domain “Golden Age” comics (c. 1940–1962) and provides reproducible benchmarks for both human and model performance in the challenging domain of comic scene segmentation (Paval et al., 22 Aug 2025).

1. Motivation and Context

NLP has long addressed text segmentation into coherent semantic units such as chapters or topics, while Computer Vision (CV) has pursued shot and scene segmentation for video data (Rao et al. 2020). However, comics combine text and image in a unique, inherently multimodal, frame-based narrative structure, yet lacked datasets finer than page- or panel-level splits. ComicScene154 conceptualizes a “scene” analogously to a narrative arc, defined as a temporally and spatially coherent plot unit pursued by a consistent cast of characters [Cohn 2013; Rao et al. 2020], aligning comics segmentation with established approaches in other storytelling media. This affords comics research a compact, lower-bandwidth abstraction comparable to video or film scenarios.

2. Data Sources and Composition

ComicScene154 draws exclusively from the Comic Book Plus archive, selecting four public-domain magazines to maximize genre diversity and story completeness:

• Alley Oop (humor)
• Champ Comics (superheroes)
• Treasure Comics (fantasy)
• Western Love (romance)

These titles comprise 34 stories across 154 pages. Each page underwent OCR cleaning; individual panels were identified with bounding-box coordinates and ordered in standard left-to-right, top-to-bottom reading sequence. The complete dataset consists of approximately 600 panels, yielding both visual and textual inputs for downstream processing tasks.

The following table summarizes dataset composition:

Magazine	Genre	Stories	Pages	Panels (~)
Alley Oop	Humor	1	—	—
Champ Comics	Superheroes	24	—	—
Treasure Comics	Fantasy	6	—	—
Western Love	Romance	4	—	—
Total	—	34	154	~600

The genre distribution encompasses 1 humor, 24 superhero, 6 fantasy, and 4 romance stories. Approximately 190 scene-start labels are present (average scene length ≈ 3 panels).

3. Annotation Protocol and Schema

Within ComicScene154, annotations are rendered at the individual-panel level. Annotators labeled each panel $s_i$ with $y_i \in \{0,1\}$, indicating whether $s_i$ initiates a new scene. The annotation task drew on Cohn's theory of narrative arc and Rao et al.'s operational framing of a scene as a “plot-based semantic unit in which an overarching task is pursued by a certain cast of characters.” No sub-arc classifications (e.g., “rising action,” “climax”) are included in the current release.

Six volunteer annotators, divided into three teams, independently labeled non-overlapping thirds of the dataset. Guidelines introduced the key concept of closure (Cohn 2010) and instructed annotators to mark any panel demarcating a perceived new narrative initiation.

4. Inter-Annotator Agreement and Reliability

To quantify reliability, the pk metric from text segmentation (Glavaš et al. 2016) was adapted:

$$pk = \frac{1}{N - k} \sum_{i=1}^{N-k} \delta(g(s_i, s_{i+k}), g^*(s_i, s_{i+k}))$$

where $k$ is half the average segment length (here, $k=3$ panels), $g(\cdot, \cdot)$ is an indicator for a true scene boundary, and $\delta$ is the XOR operator. Perfect alignment yields $pk=0$, total misalignment $pk=1$.

On a sample covering a third of the dataset, the average annotator-to-gold pk was 0.17, indicating moderate agreement given the inherent subjectivity. Scores for six random excerpts ranged from 0.06 to 0.37, reflecting individual variation in narrative segmentation.

5. Baseline Scene-Segmentation Pipeline

Rather than implementing a CNN-Transformer architecture, the authors provided a baseline leveraging a two-stage, zero-shot LLM pipeline using Google’s gemini-2.0-flash-thinking-exp. In Stage 1 (multi-modal inference), the page image, per-panel bounding box coordinates, transcribed text, and reading order index were concatenated as input. The LLM was prompted to (a) summarize narrative arcs, and (b) list indices marking new arc initiations. In Stage 2 (reasoning-based refinement), the LLM’s initial boundary predictions were reintroduced with explicit instructions to amend erroneous or missed boundaries.

A hypothetical binary-cross-entropy objective,

$y_i \in \{0,1\}$0

where $y_i \in \{0,1\}$1 is the model's predicted boundary probability for panel $y_i \in \{0,1\}$2, is proposed for future trainable extensions, though no gradient-driven training was conducted in this study.

6. Performance Benchmarking

Pipeline evaluation again employed the pk metric. In Stage 1, ten prompt runs yielded an average $y_i \in \{0,1\}$3, only marginally better than a random baseline ($y_i \in \{0,1\}$4) or an “in-between” baseline ($y_i \in \{0,1\}$5). Stage 2 refinement, over 100 LLM runs (10 initial × 10 refined), improved best-case $y_i \in \{0,1\}$6 to 0.34 but the mean remained at $y_i \in \{0,1\}$7. Human annotators, by comparison, achieved $y_i \in \{0,1\}$8. These results highlight persistent challenges for LLMs in multimodal narrative segmentation and underscore the subjectivity of the task.

7. Applications, Limitations, and Future Directions

ComicScene154 enables a range of multimodal narrative research tasks, including:

• Scene-level narrative summarization (Huang et al. 2016)
• Fine-grained entity and character tracking (Kim & Schuster 2023)
• Digital-humanities genre analysis
• Comicsification—deriving low-bandwidth scene abstractions for video or film (Rao et al. 2020)

The principal limitation is the subjectivity of scene boundary identification: even trained annotators manifest substantial divergence, as reflected in variable pk scores. Reliance on “Golden Age” Western comics constrains generalizability for modern or non-Western comic forms. Future work could expand annotation to encompass sub-arc structure (e.g., “setup,” “climax”), diversify source material, and develop fully supervised architectures that combine visual CNN-derived features $y_i \in \{0,1\}$9 and textual embeddings $s_i$0 using

$s_i$1

trained under the cross-entropy loss.

ComicScene154 fills a previously unaddressed gap in multimodal narrative analysis by delivering reproducible scene-segmented annotations, reliability benchmarks, and a transparent LLM pipeline. It lays a foundation for more robust story-understanding systems at the intersection of NLP and CV (Paval et al., 22 Aug 2025).