AniLines - Anime Lineart Extractor

©Oniichan wa Oshimai!

⭐ AniLines is a simple tool to extract lineart from anime images and videos.

🤗HuggingFace Demo

Contents

• Features
• How to use
  • Environment
  • Run Inference
  • Inference acceleration
• Gallery
• Related Works
• Acknowledgements
• Citation
• License

Updates

• Huggingface space is back online again - 2025/09/18
• Init Repo - 2025/02/10
• Support for video - 2025/02/11
• Gradio UI and 🤗Huggingface Demo - 2025/02/19

Features

What does AniLines do?

Like 🚀MangaLineExtraction and 👓Anime2Sketch, AniLines generates sketch from anime(🎞️cel animation🎞️ in particular) images, it generally provides more details and clearer lines with fewer artifacts on characters.

Input	MangaLine	Anime2Sketch	DoG(Binary)	AniLines(Basic)	AniLines(Detail)

©Urusei Yatsura 2022, ©Violet Evergarden, ©Miss Kobayashi's Dragon Maid, ©Little Witch Academia

All input images shown above were inferenced at a 1080p resolution. It is also advised to use 1080p images for better performance.

How does AniLines work?

AniLines has two modes: basic and detail. While basic model extracts the main structure of a drawing, detail model provides the sketch on more elements like backgrounds and cel edges. Personally speaking, detail mode has better performance as it has a higher 'recall' score, and consequently miss less lines.

input	basic mode	detail mode

©Watashi ni Tenshi ga Maiorita!

How to use

Environment

First, clone this repo, create a new conda environment and install the requirements:

git clone https://github.com/zhenglinpan/AniLines-Anime-Line-Extractor.git
cd AniLines-Anime-Line-Extractor

conda create -n anilines python=3.12 -y
conda activate anilines

pip install -r requirements.txt

Then, download the pre-trained models from the links below and put them in the ./weights folder

• Basic model
• Detail model

Run inference

There are two ways to use AniLines:

👉 1. Command line

python infer.py --dir_in ./input --dir_out ./output --mode detail --binarize -1 --fp16 True --device cuda:0

You can either pass a single image or a folder to dir_in, the extracted lineart will be saved to the output folder by default.

👉 2. Gradio UI

An gradio interface for AniLines is available!

Here is a demo on 🤗HuggingFace Space.

Compared with the demo, local UI allows you to process video and batch files, and presumably faster if you have better machines. It is much recommended to run locally with your gpu machine, you can simply run the following command to start the UI locally:

python app.py

Video support

AniLines supports video processing. When a video file is provided, it will generate a video of the extracted lineart.

Binarized output

Binarized lines are often used in animation production. AniLines provides an API for this feature. By default, binarization is disabled (set to -1). You can enable binarization by setting the --binarize parameter to any value between 0 and 1. This will adjust the threshold for binarizing the output.

input	no binarize	binarize 0.5	binarize 0.95

©RainbowSea

Inference acceleration

By default, AniLines uses fp16 to accelerate inference, trading marginal performance for a considerable speed boost. You can set --fp16 to False if your GPU does not support it or if you prefer full precision. When using fp16, the model runs up to 5 times faster, as tested with 1080p images.

For further accleration, you can play with TensorRT/TorchTRT with onnx to achieve its maximum potential speed.

Gallery

Here are some more examples of AniLines inference results. AniLines is basically designed to work for cel animation images, but it does maintain an acceptable transferability to other types of images, like illustrations or flat color without clear edges. This gallery will be updated with more examples in the future.

input	detail mode

©YuZhu

Related works

Here are some other anime lineart extraction projects that you might find interesting:

• MangaLineExtraction
• Anime2Sketch
• LineDistiller
• sketchKeras
• DoG(with fine-tuned hyperparameters)
• sketch-extraction-of-anime-character

Acknowledgements

Special thanks to Shuhong for his advice in this project.

Citation

Feel free to cite this work if you find it useful for your research:

@misc{AniLines,
  author = {Zhenglin Pan},
  title = {AniLines - Anime Lineart Extractor},
  publisher = {GitHub},
  year = {2025},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/zhenglinpan/AniLines-Anime-Lineart-Extractor}}
}

License

This project is licensed under the MIT License - see the LICENSE file for details.