Anime Style Transfer

Thank you @AK

Anime Style Transfer with Pytorch

• YOLOv5
• Waifu2x
• CLIP
• StyleGAN2(Blending)
• pix2pixHD

Web Demo

Integrated into Huggingface Spaces 🤗 using Gradio. Try out the Web Demo:

My Env

• python 3.7
• pytorch 1.8.1
• cuda 11.1
• cudnn 8.0.5

---

collect video and image
->
face detection
->
image filtering using clip (clear / blurred)
->
waifu upsampling
->
face alignment according to ffhq standard
->
train Anime StyleGAN (transfer learning!!)
->
FFHQ, Anime StyleGAN Blending
->
Make Dataset
->
Pix2PixHD Training
->
Result !!

---

1. Prepare custom data

# video

ffmpeg -i video.mp4 -filter:v fps=0.5 video%d.jpg

# folder
anime
  |- 1.png
  |- 2.png
  |- 3.png
  |- ...

2. Face Detection with YOLOv5

• YOLOv5

Custom Training with icartoonface dataset

• iCartoon

• icartoonface format -> yolov5 format

  • Reference : icartoonface_convert.ipynb

• train yolov5

• detect + crop

count = 0
for name in tqdm(os.listdir(folder)):
    img_path = os.path.join(folder, name)
    img = cv2.imread(img_path)
    img_h, img_w, _= img.shape
    
    results = model(img_path)
    
    for rst in results.crop():

        xmin, ymin, xmax, ymax = rst['box']

        xmin = int(xmin.cpu().numpy() * 0.92)
        xmax = int(xmax.cpu().numpy() * 1.08)
        ymin = int(ymin.cpu().numpy() * 0.92)
        ymax = int(ymax.cpu().numpy() * 1.08)

        if (xmax - xmin) > (ymax - ymin):
            ymax += (((xmax - xmin) - (ymax - ymin)) // 2)
            ymin -= (((xmax - xmin) - (ymax - ymin)) // 2)
        else:
            xmax += (((ymax - ymin) - (xmax - xmin)) // 2)
            xmin -= (((ymax - ymin) - (xmax - xmin)) // 2)

        xmax = min(xmax, img_w) 
        ymax = min(ymax, img_h) 
        xmin = max(xmin, 0) 
        ymin = max(ymin, 0) 

        cv2.imwrite(f'./result/{count}.png', img[ymin:ymax, xmin:xmax])
        count += 1

3. Remove Blurred Images with CLIP

• CLIP

["clear face", "blurred face"]

4. Image Super-Resoultion with Waifu2x

• Waifu2x

5. Face Alignment according to FFHQ standard

• FFHQ

6. Train StyleGAN2

• StyleGAN2

python dataset_tool.py --source=[YOUR DATASET PATH] --dest=./anime.zip --width 512 --height 512

python train.py --outdir=./training-runs --data=./anime.zip --cfg=paper512 --mirror=1 --gpus=4 --batch 8 --resume ffhq512

7. Blend Model

• StyleGAN2

• Reference : ./blending.ipynb

• Number of Images : 10000

8. Pix2Pix Training

• Pix2PixHD

# train

python train.py --name anime --dataroot ./datasets/anime --loadSize 512 --label_nc 0 --no_instance


# test

python test.py --name anime --dataroot ./datasets/test --loadSize 512 --label_nc 0 --no_instance

Reference

• https://github.com/bryandlee/animegan2-pytorch
• https://github.com/Sxela/ArcaneGAN