Comic Analysis: Project Overview

This document outlines the ongoing work in analyzing comics using data science techniques, focusing on multimodal modeling for content understanding and querying.

Models Available

Cosmo Page Stream Segmentation - https://huggingface.co/RichardScottOZ/cosmo-v4

• trained on a 25K page subset of more modern comics from multi companies and multiple genres and styles (not manga) Closure Lite Simple - https://huggingface.co/RichardScottOZ/comics-analysis-closure-lite-simple
• composition and structural analysis

1. Project Goals & Evolution

• Version 1 (Closure Lite Framework):
  • Aim: To combine page image and VLM text for querying, with panel and reading order recognition.
  • Key Choice: Utilized 384-dimensional "lite" embeddings to allow for testing on retail GPUs.
  • Model: Based on the RichardScottOZ/comics-analysis-closure-lite-simple model, trained as a context-denoise variant.
• Version 2 (Current Focus): Initiating work towards a new, more robust modeling framework, particularly for Page Stream Segmentation.

2. Core Framework: Closure Lite

The Closure Lite Framework is a multimodal fusion model designed to analyze images, text, and panel reading order.

• Objective: Generate multimodal fused embeddings that are queryable for similarity.
  • Embeddings are efficiently stored in Zarr format (e.g., 80K "perfect match subset" takes ~500MB).
  • An interface (Flask-based) exists for querying.
• Design Principle: Usable on a reasonable retail GPU (hence 384-dim embeddings).
• Model Variants: base, denoise, context, context-denoise.

3. Data Sources & Preparation

• Primary Data: Comic pages from various sources (Amazon/Comixology, Dark Horse, Calibre-managed collections like Humble Bundle and DriveThruComics).
• Future Data: Exploring new Humble Bundles and potential scraping (e.g., Neon Ichigan).
• Initial Processing:
  • Conversion: Comic files (PDF, CBZ, etc.) are converted into individual page images.
  • VLM Text Extraction: Vision-Language Models (VLMs) are used to extract text from each page, preceding general OCR due to its comic-specific challenges.
  • Panel Detection: Fast-RCNN is used to detect panel bounding boxes and coordinates.
  • DataSpec Integration: All extracted data is joined into a DataSpec format for modeling.
• Dataset Size: Currently working with a "Calibre" subset of ~330K pages (compared to ~805K Amazon pages).

4. Basic Workflow

1. Find Comics: Gather digital comic files.
2. Convert to Pages: Extract images from containers (PDF, CBZ, etc.).
3. VLM Analysis: Use VLMs to get panel text for each page.
4. R-CNN Panel Detection: Get panel bounding boxes and coordinates.
5. Data Integration: Join data into DataSpec (e.g., using coco_to_dataspec).
6. Perfect Match Subset: Create high-quality subsets for training.
7. Model Training: Train the Closure Lite model (using closure_lite_dataset, closure_lite_simple_framework, train_closure_lite_simple_with_list).
8. Embedding Generation: Run the trained model to generate queryable embeddings.
9. Querying: Access embeddings via code or a dedicated interface.

5. Perfect Match Dataset Notes

• Goal: Ensure text alignment with the correct panels (same number of panels from R-CNN and VLM).
• Challenge: Initial alignment between Fast-RCNN and various VLM runs was only ~25%, indicating a need to improve this step or consider alternative approaches (e.g., OCR for Fast-RCNN).

6. Key Results from Closure Lite (80K Calibre Subset, Context-Denoise)

Panel Similarity

• An early model showed panel similarities, with the model cutting things off at 12 panels.

Page Similarity

• The model demonstrated understanding of page composition and structure.

Clustering (UMAP Embeddings)

• UMAP embeddings revealed distinct clusters.
• Landscape Cluster: A "ribbon cluster" was identified as a "landscape cluster" (pages with landscape orientation), which is less common in comics.
• Data Noise: An adjoining part of this cluster was found to be an error, containing extracted pages from non-comic content (e.g., computing books), highlighting the need for better data filtering.

Ablation Study (Vision-Only Embeddings)

• Clustering of vision-only embeddings showed distinct patterns.
• Cluster 8: Identified as "all single panel very dark images" and "heavy text one page images," with a few multi-panel pages. A disconnected group of "all white images" formed a monochromatic sub-cluster.
• Aspect Ratios: Vision-only clustering also picked out significant differences in panel aspect ratios.

7. Relevant Research & Future Tasks

Relevant Research

• Survey: "One missing piece in Vision and Language: A Survey on Comics Understanding" (https://arxiv.org/abs/2409.09502v1)

• Neural Networks & Transformers: "Investigating Neural Networks and Transformer Models for Enhanced Comic Decoding" (https://dl.acm.org/doi/10.1007/978-3-031-70645-5_10)

• Image Indexing: "Digital Comics Image Indexing Based on Deep Learning" (https://www.researchgate.net/publication/326137469_Digital_Comics_Image_Indexing_Based_on_Deep_Learning)

• Synthesis of Graphic Novels: "A Deep Learning Pipeline for the Synthesis of Graphic Novels" (https://computationalcreativity.net/iccc21/wp-content/uploads/2021/09/ICCC_2021_paper_52.pdf)

• ComicsPAP: https://github.com/llabres/ComicsPAP =======

• Papers:

  • https://arxiv.org/abs/2503.08561

  • https://www.researchgate.net/publication/389748978_ComicsPAP_understanding_comic_strips_by_picking_the_correct_panel? _tp=eyJjb250ZXh0Ijp7ImZpcnN0UGFnZSI6InB1YmxpY2F0aW9uIiwicGFnZSI6InB1YmxpY2F0aW9uIn19

    • https://huggingface.co/VLR- ComicsPap models
    • https://github.com/emanuelevivoli/CoMix - CoMix repo

  • https://www.researchgate.net/publication/326137469_Digital_Comics_Image_Indexing_Based_on_Deep_Learning

  • https://computationalcreativity.net/iccc21/wp-content/uploads/2021/09/ICCC_2021_paper_52.pdf

Tasks of Interest

• Dialogue transcription, which necessitates a robust panel detection pipeline.

8. Pipeline Details

• Refer to the main CoMix repository for the Faster RCNN panel detection and processing to make coco style fiels
• The code in detections_2000ad contains adaptations for inference, distinct from evaluation.
• Detailed functions are in https://github.com/RichardScottOZ/CoMix/tree/main/benchmarks/detections.
• Technical Stack: PyTorch (CUDA 11.8 recommended), transformers, xarray and zarr.

9. VLM Model Experiments & Possibilities

Extensive testing has been conducted with various Vision-Language Models for text extraction and analysis:

• Gemini 4B: Handles basics, but fails on some perpetually (reason yet to be understood; Fourier analysis suggested).
• Gemma 12B: Can handle some failures of Gemini 4B.
• Mistral 3.1: Generally identifies "character A and character B," can handle Gemini failures.
• Qwen 2.5 VL Instruct: Failed on the same cases as Gemini 4B.
• Phi4: Not very good.
• Llama 11B: Failed to process.
• Gemini Flash 1.5: Capable of handling missing/null captions and characters.
• Gemini Flash 2.5 Lite: Good performance, but 8x more expensive than Gemma 4B (which can run locally). Experienced quite a few connection errors with Google. On the hardest 660 samples, had 1/3 errors and 1/5 JSON errors.
• GPT Nano 4.1: Reported "unsupported image type," indicating lower capability than Google's models.
• Meta Llama 4 Scout: Much better, with 300/500 success on remaining images, and better cost efficiency (0.08/0.3 vs 0.10/0.40 for Gemini Flash Lite 2.5). However, encountered significant problems with the GMI Cloud provider.

10. Embeddings Generation Strategy

• To Generate: Panel embeddings (P), Page embeddings (E_page), Reading order embeddings.
• Dataset Coverage: Amazon perfect matches (212K pages), CalibreComics perfect matches (80K), and a combined dataset of all high-quality samples.
• Technical Approach: Batch Processing Script.

11. Domain capability

• Humble Bundle: If you download everything you will get your book and rpg collections too - which can happily go through pdf and epub extractions and end up with images.

• Framework handled this, the 'deep learning / linux / programming' book images clusters as their own small part connected to a ribbon, so not ideal to include but a useful result. So what if you put rpg art in? Lots of single page 'panels' obtained most likely - and would cluster in the 'no/low test text' UMAP sections

12. Future Research

• Refer to documentation/future_work/.

Version 2

1. Paratext

• Need to work out comic classes inside a comic that will work for 'modern comics' as opposed to old public domain comics
• Cover, credits and indicia, splash, cover interior for trades, back matter text, back matter art, previews etc.

Problems

• single issues
• trade paperbacks
• omnibuses, collections that are larger
• anthologies - art/narrative changes 'inside' a comic - e.g. 2000AD
  • anthologies inside omnibuses

a. Zero shot classification

• Trying bart for the above as have many VLM json files with some signal of what pages might include in 'overall summary' for example.

b. Page Stream Segmentation (PSS) & CoSMo

• Importance: PSS is crucial for the next version of the project, feeding page type markers into multimodal fusion.
• CoSMo Model:
  • GitHub Repository (includes arXiv paper link).
  • Specifically designed for PSS.
  • Uses Qwen 2.5 VL 32B for OCR.
  • Consideration: Is Gemma as good/cost-effective for OCR?
  • Current Status: No pre-trained model available for direct testing; requires training a new one.

• notebook in Data Analysis - comic training
  • try first pass classes
  • cover, credits, advertisements, art (previews, internal covers etc. hopefully eventually)), text, back_cover - which could be ads or previews often so might not be a useful classification

2. OCR

• Needed for everything - page / panel / characters
• Discussion on Deepseek OCR references Qwen here too: https://news.ycombinator.com/item?id=43549072
  • note bounding box commentary
  • also https://pyimagesearch.com/2025/06/09/object-detection-and-visual-grounding-with-qwen-2-5/