ToonOut Dataset

Please check out:

• our repository: https://github.com/MatteoKartoon/BiRefNet
• our paper: ToonOut: Fine-tuned Background Removal for Anime Characters
• the weights for BiRefNet, fine-tuned on this dataset and the associated Replicate endpoint if you want to try out the model (courtesy of J. H. Cho @sprited)

Dataset Summary

The ToonOut Dataset is a collection of 1,228 high-quality anime-style images annotated for background removal tasks. Each sample includes raw RGB images, ground truth transparency masks, and RGBA annotated images with alpha channels. The dataset was specifically curated to improve segmentation and matting performance on anime characters and objects, addressing challenges such as complex hair, transparency, and stylized features.

It was used to fine-tune the BiRefNet model, resulting in ToonOut, which significantly improves background removal accuracy for anime-style images.

• License: CC-BY 4.0 (dataset)
• Intended task: Image segmentation (foreground/background separation for anime characters)
• Size: 1,228 annotated images (with train/val/test splits)

---

Dataset Details

Dataset Description

• Curated by: Matteo Muratori, Joël Seytre, Gabriel Aubourg
• Funded by: Kartoon AI
• Shared by: Matteo Muratori & Joël Seytre
• Language(s): Not applicable (image dataset)
• License: CC-BY 4.0

---

Uses

Direct Use

• Training, fine-tuning, or benchmarking background removal models for anime-style content. .

---

Sample Usage

To get started with the associated code and dataset, follow the installation steps and then use the provided scripts for training and evaluation.

Installation

git clone https://github.com/MatteoKartoon/BiRefNet.git
cd BiRefNet
pip install -r requirements.txt

Inference demo

Please see the provided notebook here.

Training

bash bash_scripts/train_finetuning.sh

Evaluation

python scripts/evaluations.py --checkpoint path/to/checkpoint

---

Dataset Structure

The dataset is split into train, validation, and test sets. Each split is organized into generation folders containing three subfolders:

• im/: Raw RGB images
• gt/: Ground truth transparency masks
• an/: RGBA annotated images (RGB + alpha)

toonout_dataset/
├── train/
│   ├── train_generations_20250318_emotion/
│   │   ├── im/    # RGB images
│   │   ├── gt/    # Ground truth masks
│   │   └── an/    # RGBA annotated images
│   └── train_generations_/…
├── test/
│   └── test_generations_/…
└── val/
└── validation_generations_*/…

---

Dataset Creation

Curation Rationale

State-of-the-art background removal models underperform on stylized anime imagery, particularly with fine-grained details such as hair strands, line art, and semi-transparent effects. This dataset was curated to bridge that gap and provide training material for anime-specific background-removal models.

Source Data

Data Collection and Processing

• Images were curated from synthetic generations of anime-style images (Stable Difussion XL, yamer's anime checkpoint).
• Expert annotators manually produced transparency masks for each image, starting from a model-generated annotation.
• Splits (train/val/test) were created with care to avoid overlap.

Who are the source data producers?

• The dataset was created and annotated by researchers at Kartoon AI and collaborators.

Annotations

Annotation Process

• Annotation type: Foreground/background transparency masks.
• Tools used: Manual tools for annotation of transparency masks.
• Quality control: Manual inspection and consistency checks across the dataset.

Who are the annotators?

• Researchers at Kartoon AI curated and validated the annotations.

Personal and Sensitive Information

• The dataset contains no personal, private, or sensitive information. It consists exclusively of synthetic images.

---

Citation

If you use this dataset, please cite the accompanying paper:

BibTeX:

@misc{muratori2025toonout,
  title={ToonOut: Fine-tuned Background Removal for Anime Characters},
  author={Muratori, Matteo and Seytre, Joël},
  year={2025},
  eprint={2509.06839},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2509.06839},
  doi={10.48550/arXiv.2509.06839}
}

⸻

Dataset Card Authors: Joël Seytre (Kartoon AI)

Dataset Card Contact: Joël Seytre (joel@kartoon.ai), Matteo Muratori (matteo.muratori9@studio.unibo.it)

⸻

Project by Kartoon AI, powering toongether, check us out at kartoon.ai & toongether.ai