This repo contains a PyTorch from-scratch implementation of the paper Towards the Automatic Anime Characters Creation with Generative Adversarial Networks.

Generating 二次元(にじげん) MOEst avatars...

• Python 3.6
• Python-OpenCV 3.4.0
• pyquery 1.2.4
• i2v 1.0.0
• PyTorch 0.4.0
• GPU is better :-D
• and so on.

I want to call it a DRAGAN-like SRGAN structure because I use the gradient penalty as the paper told and two SRResNet as discriminator and generator. The SRResNet(modified as the paper described) are like this:

I have some modifications in this structure:

1. I weighted label's loss and tag's loss with half of λ_adv beacause the loss described in the paper was so hard-core for me. (More details please refer in src/model/gan.py)

2. Remove the sigmoid operation in adversarial loss calculating since the results with sigmoid layer may cause some problems.

3. Using Multi-Label Soft Margin Loss for tags' loss calculating.(Cross Entropy Loss may better because of the imbalance of the images' tag distribution. But I don't have too much time for weights tuning. :-D )

0. Cause I built an extremely clean dataset for this task. I'm glad to share my data-Preparing method here.

1. Crawled the images from the website as the paper proposed. Read the readme and codes in src/dataset/Spider/ to get more information.

2. I used the lbpcascade_animeface model for face detecting. source codes are in src/dataset/FaceDetect/

3. illustration2vec was used for face tagging. Please check the files in src/tag/

4. Remove the invalid images manually.

1. The discriminator and generator were defined in src/model/networks/.
2. The training strategy of GAN was written in src/model/gan.py.

1. Currently I have some simple tools for face data statistics. More details in src/statistics/dataset/.

Coming soon.