DACoN: DINO for Anime Paint Bucket Colorization with Any Number of Reference Images (ICCV 2025)

Usage

1. Setup
2. Train
3. Test
4. Inference

Setup

Clone the repository:

git clone https://github.com/kzmngt/DACoN.git

Move to the project directory:

cd DACoN

To setup the environment, run:

conda create -n dacon python=3.10
conda activate dacon
pip install -r requirements.txt

Train

First, you have to download the dataset from BasicPBC. Place the dataset as follows:

├──dataset
    ├──PaintBucket_Char_v2/
        ├──test/
        ├──train/

Once dowloading datasests is finished, run the following command to train DACoN from scratch:

python dacon/train.py \
    --config configs/train.yaml \
    --version 1_1

The checkpoints are saved under:

├──checkpoints
    ├──dacon_<STARTTIME>.pth
    ├──model_epoch_1.pth
    ├──model_epoch_2.pth
    ├──...

Note:

• If model_epoch_*.pth files exist in this folder, training will resume from the latest checkpoint instead of starting from scratch.

Test

You can download the pretrained checkpoints from the following links and puts under checkpoints folder.

Model	Download	Description
DACoN v1.0	link	Same architecture as paper
DACoN v1.1	link	Set a fixed image size during segment pooling. Fuse DINO and U-Net features using simple addition instead of concat + MLP.

You can evaluate model using the following command:

python dacon/test.py \
    --config configs/test.yaml \
    --model checkpoints/dacon_v1_1.pth \
    --version 1_1

You can control the following behaviors via the YAML configuration (test.yaml):

• Keyframe or Consecutive-frame colorization
• Number of reference images used for keyframe colorization

Inference

You can also colorize your own anime line images. Please organize your data as follows:

├──inference
    ├──character_A
    	├──line
            ├──0.png	
            ├──1.png	
            ├──...	
    	├──ref
            ├──gt
            	├──a.png
            	├──b.png
            	├──....
            ├──line
            	├──a.png
            	├──b.png
            	├──....
    ├──character_B
    	├──line
    	├──ref

Note:

• Only fully enclosed line images can be colorized properly. If there are gaps in the lines, the colorization quality may be degraded.
• When using RGB line drawings, pixels with white [255, 255, 255] are treated as the background.

After preparing your data, run the following command:

python dacon/inference.py \
    --config configs/inference.yaml \
    --model checkpoints/dacon_v1_1.pth \
    --data inference \
    --version 1_1

The colorized results can be found in the pred folder within each character's directory.

Citation

If you find this work useful, please cite our paper:

@inproceedings{nagata2025dacon,
  title     = {{DACoN}: {DINO} for Anime Paint Bucket Colorization with Any Number of Reference Images},
  author    = {Nagata, Kazuma and Kaneko, Naoshi},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  year      = {2025}
}