Note: I have restructured the whole folder, cleaned up training code, pruned my dataset, and updated most of the results. You can see the old version of this repo in old.

Generate colorful anime characters using GAN.

By interpolating generator input, we can see some interesting results.

By fixing color classes and varying noise, we can generate anime characters with same colors but different identity.

I noticed that anime produced in different years have distinctive artistic styles. To name a few:

• Characters in older anime tend to have larger, angular eyes, while newer anime have characters with round eyes.
• Colors are brighter and more saturated in older anime, while newer anime have mild colors.

By conditioning the ACGAN on year labels, we can generate characters with different artistic styles.

Interpolating along the year latent code:

Modify config.yaml as you wish.

> python3 run.py

The current dataset is a composition of 2 datasets:

1. One dataset with eye and hair color labels (30k+ images)
2. One dataset with year label, which is the year in which the anime is produced (60k+ images).

The dataset format is as follows:

- images/
    - XXXXX.jpg
    - ...
- labels.pkl
- eye_label.json
- year_label.json
- hair_label.json

After loading in labels.pkl with pickle, you will get a dictionary of { filname : labels }. The labels are formatted as (eye, hair, year) tuples.

{
    "32455": (8, 10, 5),
    ...
}

This means 32455.jpg has eye class 8, hair class 10, year class 5.

Missing labels will be a None. All images from dataset 1 will have year labels None, while all images from dataset 2 will have eye and hair label None.

Source code in the current repo is used to train on the first dataset. This requires some manual preprocessing (see dataset/anime_dataset.py) to extract the first dataset from the whole dataset.

The .json files map discrete labels to semantics.

// eye_label.json
{
    "aqua": 0, 
    "black": 1, 
    ...
}

• When training on the first dataset, adding some color transformations to images as a preprocessing step traning might help. You can achieve this through various torchvisions.transforms.functional methods.
• Train with N(0, 1) but sample from Gaussian of smaller variance when evaluating. Just an engineering hack to get better results.