ComicsPAP-Qwen2.5-VL: Sequential Narrative Understanding

Fine-tuning of Qwen2.5-VL-7B on the ComicsPAP benchmark for the "Pick A Panel" task. Achieved 66.41% Validation Accuracy, significantly outperforming zero-shot baselines.

Repository Structure

• src/train_qlora.py: Training script using PEFT and QLoRA.
• src/evaluate.py: Evaluation pipeline on Val/Test splits.
• official_repo/: Integrated utilities from the official ComicsPAP repository.
  • data_utils.py: Contains the SingleImagePickAPanel processor used for training.
• slurm/: Job scripts for the Snellius Supercomputer.

Project Highlights

• Model: Qwen2.5-VL-7B-Instruct (Vision-Language).
• Technique: QLoRA (Rank 16, Alpha 32) for parameter-efficient fine-tuning.
• Compute: Trained on 1x NVIDIA A100 (80GB) on the Snellius Supercomputer (SURFcua - TU/e).
• Performance:
  • Ours (7B Fine-tuned): 66.41%
  • Base 72B Zero-shot: 46.88% (from paper)
  • Base 7B Zero-shot: 30.53% (from paper)

Getting Started

1. Installation

git clone [https://github.com/kaj04/ComicsPAP-Project.git](https://github.com/kaj04/ComicsPAP-Project.git)
cd NOME_REPO
pip install -r requirements.txt

2. Run Evaluation

You can run the evaluation using my pre-trained adapters from Hugging Face:

python src/evaluate.py --adapter_path kaj04/Qwen2.5-VL-7B-ComicsPAP-QLoRA --split val

## Acknowledgments
This project is based on the **ComicsPAP** dataset and benchmark. Special thanks to the authors for their work in visual narrative understanding. 

- **Official Dataset:** [VLR-CVC/ComicsPAP](https://huggingface.co/datasets/VLR-CVC/ComicsPAP)
- **Official Paper:** [ComicsPAP: Understanding Comic Strips by Picking the Correct Panel](https://arxiv.org/abs/2503.08561)

Citation

@InProceedings{vivoli2025comicspap,
  author="Vivoli, Emanuele and Llabr{\'e}s, Artemis and Souibgui, Mohamed Ali and Bertini, Marco and Llobet, Ernest Valveny and Karatzas, Dimosthenis",
  editor="Yin, Xu-Cheng and Karatzas, Dimosthenis and Lopresti, Daniel",
  title="ComicsPAP: Understanding Comic Strips by Picking the Correct Panel",
  booktitle="Document Analysis and Recognition -- ICDAR 2025",
  year="2026",
  publisher="Springer Nature Switzerland",
  address="Cham",
  pages="337--350",
  isbn="978-3-032-04614-7"
}