PyTorch Implementation of AnimeGANv2

Updates

• 2021-10-17 Add weights for FacePortraitV2.

  

• 2021-11-07 Thanks to ak92501, a web demo is integrated to Huggingface Spaces with Gradio.

• 2021-11-07 Thanks to xhlulu, the torch.hub model is now available. See Torch Hub Usage.

Basic Usage

Inference

python test.py --input_dir [image_folder_path] --device [cpu/cuda]

Torch Hub Usage

You can load the model via torch.hub:

import torch
model = torch.hub.load("bryandlee/animegan2-pytorch", "generator").eval()
out = model(img_tensor)  # BCHW tensor

Currently, the following pretrained shorthands are available:

model = torch.hub.load("bryandlee/animegan2-pytorch:main", "generator", pretrained="celeba_distill")
model = torch.hub.load("bryandlee/animegan2-pytorch:main", "generator", pretrained="face_paint_512_v1")
model = torch.hub.load("bryandlee/animegan2-pytorch:main", "generator", pretrained="face_paint_512_v2")
model = torch.hub.load("bryandlee/animegan2-pytorch:main", "generator", pretrained="paprika")

You can also load the face2paint util function:

from PIL import Image

face2paint = torch.hub.load("bryandlee/animegan2-pytorch:main", "face2paint", size=512)

img = Image.open(...).convert("RGB")
out = face2paint(model, img)

More details about torch.hub is in the torch docs

Weight Conversion from the Original Repo (Tensorflow)

1. Install the original repo's dependencies: python 3.6, tensorflow 1.15.0-gpu
2. Install torch >= 1.7.1
3. Clone the original repo & run

git clone https://github.com/TachibanaYoshino/AnimeGANv2
python convert_weights.py

samples

Results from converted `Paprika` style model (input image, original tensorflow result, pytorch result from left to right)

     

Note: Results from converted weights slightly different due to the bilinear upsample issue

Additional Model Weights

Webtoon Face [ckpt]

samples

Trained on 256x256 face images. Distilled from webtoon face model with L2 + VGG + GAN Loss and CelebA-HQ images.

 

Face Portrait v1 [ckpt]

samples

Trained on 512x512 face images.

📺

Face Portrait v2 [ckpt]

samples

Trained on 512x512 face images. Compared to v1, 🔻beautify 🔺robustness

🦑 🎮 🔥