This repository contains a PyTorch implementation of the following paper:
A Style-Based Generator Architecture for Generative Adversarial Networks
Tero Karras (NVIDIA), Samuli Laine (NVIDIA), Timo Aila (NVIDIA)
http://stylegan.xyz/paperAbstract: We propose an alternative generator architecture for generative adversarial networks, borrowing from style transfer literature. The new architecture leads to an automatically learned, unsupervised separation of high-level attributes (e.g., pose and identity when trained on human faces) and stochastic variation in the generated images (e.g., freckles, hair), and it enables intuitive, scale-specific control of the synthesis. The new generator improves the state-of-the-art in terms of traditional distribution quality metrics, leads to demonstrably better interpolation properties, and also better disentangles the latent factors of variation. To quantify interpolation quality and disentanglement, we propose two new, automated methods that are applicable to any generator architecture. Finally, we introduce a new, highly varied and high-quality dataset of human faces.
Picture: These people are not real – they were produced by our generator that allows control over different aspects of the image.
To the best of my knowledge, there is still not a similar pytorch 1.0 implementation of styleGAN as NvLabs released(Tensorflow), therefore, i wanna implement it on pytorch1.0.1 to extend its usage in pytorch community.
@date: 2019.10.21
@info: The noteworthy thing I just ignore to highlight is you need to change default Star
dataset to your own dataset (such as FFHQ or others) in opts.py
. Sorry for my carelessness for this.
# ① pass your own dataset of training, batchsize and common settings in TrainOpts of `opts.py`.
# ② run train_stylegan.py
python3 train_stylegan.py
# ③ you can get intermediate pics generated by stylegenerator in `opts.det/images/`
we follow the release code of styleGAN carefully and if you found any bug or mistake in implementation, please tell us and improve it, thank u very much!
blur2d
mechanism. (a step which takes much gpu memory and if you don't have enough resouces, please set it toNone
.)truncation
tricks.- Two kind of
upsample
method inG_synthesis
. - Two kind of
downsample
method inStyleDiscriminator
. PixelNorm
andInstanceNorm
.Noise
mechanism.styleMixed
mechanism.- add
Multi-GPU
support.
- Inference code.
1. StyleGAN - Official TensorFlow Implementation
2. The re-implementation of style-based generator idea
- Ubuntu18.04
- PyTorch 1.0.1
- Numpy 1.13.3
- torchvision 0.2.1
- scikit-image 0.15.0
- tqdm
- GTX 1080Ti or above
Our code can run 1024 x 1024
resolution image generation task on 1080Ti, if you have stronger graphic card or GPU, then
you may train your model with large batchsize and self-define your multi-gpu version of this code.
My Email is [email protected], if you have any question and wanna to PR, please let me know, thank you.