Skip to content

Latest commit

 

History

History
138 lines (98 loc) · 5.26 KB

README_EN.md

File metadata and controls

138 lines (98 loc) · 5.26 KB
Raw

Photo to Cartoon

中文版 | English Version

Minivision's photo-to-cartoon translation project is opened source in this repo, you can try our WeChat mini program "AI Cartoon Show" via scanning the QR code below.

Introduce

The aim of portrait cartoon stylization is to transform real photos into cartoon images with portrait's ID information and texture details. We use Generative Adversarial Network method to realize the mapping of picture to cartoon. Considering the difficulty in obtaining paired data and the non-corresponding shape of input and output, we adopt unpaired image translation fashion.

The results of CycleGAN, a classic unpaired image translation method, often have obvious artifacts and are unstable. Recently, Kim et al. propose a novel normalization function (AdaLIN) and an attention module in paper "U-GAT-IT" and achieve exquisite selfie2anime results.

Different from the exaggerated anime style, our cartoon style is more realistic and contains unequivocal ID information. To this end, we add a Face ID Loss (cosine distance of ID features between input image and cartoon image) to reach identity invariance.

We propose a Soft Adaptive Layer-Instance Normalization (Soft-AdaLIN) method which fuses the statistics of encoding features and decoding features in de-standardization.

Based on U-GAT-IT, two hourglass modules are introduced before encoder and after decoder to improve the performance in a progressively way.

We also pre-process the data to a fixed pattern to help reduce the difficulty of optimization. For details, see below.

Start

Requirements

  • python 3.6
  • pytorch 1.4
  • tensorflow-gpu 1.14
  • face-alignment
  • dlib

Clone

git clone https://github.com/minivision-ai/photo2cartoon.git
cd ./photo2cartoon

Download

Google Drive | Baidu Cloud acess code: z2nm

  1. Put the pre-trained photo2cartoon model photo2cartoon_weights.pt into models folder.
  2. Place the head segmentation model seg_model_384.pb in utils folder.
  3. Put the pre-trained face recognition model model_mobilefacenet.pth into models folder (From InsightFace_Pytorch).
  4. Open-source cartoon dataset cartoon_data/ contains trainB and testB.

Test

Please use a young Asian woman photo.

python test.py --photo_path ./images/photo_test.jpg --save_path ./images/cartoon_result.png

Train

1.Data

Training data contains portrait photos (domain A) and cartoon images (domain B). The following process can help reduce the difficulty of optimization.

  • Detect face and its landmarks.
  • Face alignment according to landmarks.
  • expand the bbox of landmarks and crop face.
  • remove the background by semantic segment.

We provide 204 cartoon images, besides, you need to prepare about 1,000 young Asian women photos and pre-process them by following command.

python data_process.py --data_path YourPhotoFolderPath --save_path YourSaveFolderPath

The dataset directory should look like this:

├── dataset
    └── photo2cartoon
        ├── trainA
            ├── xxx.jpg
            ├── yyy.png
            └── ...
        ├── trainB
            ├── zzz.jpg
            ├── www.png
            └── ...
        ├── testA
            ├── aaa.jpg 
            ├── bbb.png
            └── ...
        └── testB
            ├── ccc.jpg 
            ├── ddd.png
            └── ...

2.Train

Train from scratch:

python train.py --dataset photo2cartoon

Load pre-trained weights:

python train.py --dataset photo2cartoon --pretrained_weights models/photo2cartoon_weights.pt

Q&A

Q：Why is the result of this project different from mini program?

A: For better performance, we customized the cartoon data (about 200 images) when training model for mini program. We also improved input size for high definition. Besides, we adopted our internal recognition model to calculate Face ID Loss which is much better than the open-sourced one used in this repo.

Q: How to select best model?

A: We trained model about 200k iterations, then selected best model according to FID metric.

Q: About face recognition model.

A: We found that the experimental result calculated Face ID Loss by our internal recognition model is much better than the open-sourced one. You can try to remove Face ID Loss if the result is unstable.

Q：Can I use the segmentation model to predict half-length portrait?

A：No. The model is trained for croped face specifically.

Reference

U-GAT-IT: Unsupervised Generative Attentional Networks with Adaptive Layer-Instance Normalization for Image-to-Image Translation [Paper][Code]

InsightFace_Pytorch