Official code repository for the paper Transforming Facial Weight of Real Images by Editing Latent Space of StyleGAN.
In case you find any of this useful, consider citing:
@ARTICLE{2020arXiv201102606R,
author = {{Pinnimty}, V N S Rama Krishna and {Zhao}, Matt and
{Achananuparp}, Palakorn and {Lim}, Ee-Peng},
title = "{Transforming Facial Weight of Real Images by Editing Latent Space of StyleGAN}",
journal = {arXiv e-prints},
keywords = {Computer Science - Computer Vision and Pattern Recognition, Computer Science - Artificial Intelligence, Computer Science - Machine Learning},
year = 2020,
month = nov,
eid = {arXiv:2011.02606},
pages = {arXiv:2011.02606},
archivePrefix = {arXiv},
eprint = {2011.02606},
primaryClass = {cs.CV},
adsurl = {https://ui.adsabs.harvard.edu/abs/2020arXiv201102606R},
adsnote = {Provided by the SAO/NASA Astrophysics Data System}
}
The structure of the code is adopted from https://github.com/pbaylies/stylegan-encoder.
Material related to our paper is available via the following links:
- Paper: https://arxiv.org/abs/2011.02606
- Code: https://github.com/ramapinnimty/weight-change-visualizer
- Demo WebApp: https://research.larc.smu.edu.sg/facialweight
- Google Drive data repository: https://drive.google.com/drive/folders/1SjpWM_7CkEroJoKhOsx7zTcg-09oPxqJ?usp=sharing
All the additional material including the original data, pre-trained models, results, etc can be found at the above Google Drive link. The folder structure is as follows:
| Path | Description |
|---|---|
| PROJECT: Facial Weight Transformation | Main folder. |
| ├ paper.pdf | High-quality version of the paper PDF. |
| ├ cache | Folder with all the pre-trained models. |
| ├ Logs | Folder with detailed log files. |
| │ ├ Face Alignment | Logging info for Step-1 i.e, Pre-processing. |
| │ ├ Image Embedding | Logging info for Step-2 i.e, Latent Space Embedding. |
| ├ Metrics | Folder with score reports for all the evaluation methodologies used in the paper. |
| │ ├ Metrics.xlsx | Spreadsheet detailing the score summary of all the evaluation strategies. |
| │ ├ Image Embedding | Metrics used to evaluate the quality of Latent Space Embedding. |
| │ ├ LPIPS | Learned Perceptual Image Patch Similarity (LPIPS) scores. |
| │ ├ PSNR | Peak signal-to-noise ratio (PSNR) scores. |
| │ ├ SSIM | Structural Similarity Index Measure (SSIM) scores. |
| │ ├ Image Transformation | Metrics used to evaluate the overall quality of Facial Weight Transformation. |
| │ ├ FID | Frechet Inception Distance (FID) scores. |
| │ ├ Openface | CMU OpenFace scores. |
| ├ Results | Results for all the performed experiments. |
| │ ├ AMT | Human evaluation results. |
| │ ├ Case Study | Results generated using CycleGAN and Pix2Pix models. |
| │ ├ Face Alignment | Cropped and aligned face images at 1024x1024 resolution. |
| │ ├ Image Embedding | Folder contains embedded images along with their corresponding latent codes. |
| │ ├ Image Transformation | Folder contains the thinnest(-5), thinner(-3), normal(0), heavy(+3), and heavier(+5) transformations. |
| ├ Test set | Original data used for all the experiments in the paper. |
| │ ├ AMT | Data used for the AMT user evaluation. |
| │ ├ Feature Vector | Data used to train the weight attribute classifier. |
| │ ├ Real Images | Folder with 228 real images i.e, 12 Celebrity images, 100 images from the CFD dataset, 16 images used in the Deep Shapely Portraits paper, and 100 images from the WIDER FACE dataset. |
| │ ├ Synthetic Images | Folder with 100 randomly generated StyleGAN images. |
- Linux with Ubuntu 18.04.3 LTS or higher.
- Docker version 19.03.4.
- 64-bit Python 3.6 installation. We recommend installation through
pipwith numpy 1.19.2 or newer. - TensorFlow 1.14.0 with GPU support.
- One or more high-end NVIDIA GPUs with at least 11GB of DRAM. We recommend NVIDIA DGX-1 with 8 Tesla V100 GPUs.
- NVIDIA driver 440.64 or newer, CUDA toolkit 10.2 or newer.
If you face any problem in running this code, you can contact us at {ramap, mattzhao}@smu.edu.sg.
Copyright (c) 2020 V N S Rama Krishna Pinnimty, Matt Zhao, Palakorn Achananuparp, Ee-Peng Lim.
For license information, see LICENSE or http://mit-license.org