Project Page | Video | Paper
MyStyle: A Personalized Generative Prior
Yotam Nitzan1,2, Kfir Aberman1, Qiurui He1, Orly Liba1, Michal Yarom1, Yossi Gandelsman1, Inbar Mosseri1, Yael Pritch1, Daniel Cohen-Or2
1 Google Research, 2 Tel-Aviv University
Code was tested with Python 3.9.5, Pytorch 1.9.0 and CUDA 11.4.
We provide a yml file to easy setup of a conda environment.
conda env create -f mystyle_env.yml
Auxilary pre-trained models are required for some workflows and are listed below.
| Name | Data Preprocessing | Training | Description |
|---|---|---|---|
| arcface | ✔️ | Face-recognition network taken from TreB1eN. | |
| dlib landmarks model | ✔️ | Used for face alignment. | |
| FFHQ StyleGAN | ✔️ | StyleGANv2 model trained on FFHQ. | |
| W Inversion Encoder | ✔️ | Taken from HyperStyle. |
We proivde a pre-processing pipeline so you can easily transform a personal photo album into a suitable training set. Start by having the photo album under a single local directory.
Note: Interactive scripts below require GUI / X-11.
We first crop and align the faces in the photos. Naturally, besides the intended person, faces of other people may appear in the photo album. To filter them out run:
python data_preprocess/0_align_face.py \
--images_dir /path/to/photo/album \
--save_dir /path/to/save/resulting/images \
--trash_dir /path/to/save/unused/images \
--landmarks_model /path/to/dlib/landmarks/model \
--id_model /path/to/arcface/model \
To learn who is the intended person, you will be prompted a few images of faces. If the face belongs to the intended person press Y to confirm. Press any other key otherwise. After having enough positive samples, we'll filter the rest of the images automatically.
If you are certain the photo album contains only the face of the desired person, you do not need to pass --id_model and the alignment is fully automatic.
From the resulting aligned set, we filter out unsuitable images if they are either low quality or grayscale. Run:
python data_preprocess/1_image_quality.py \
--images_dir /path/to/aligned/faces \
--save_dir /path/to/save/resulting/images \
--trash_dir /path/to/save/unused/images
Last, we'll remove duplicates from our set. To do that, run:
python data_preprocess/2_duplicates.py \
--images_dir /path/to/current/filtered/images \
--trash_dir /path/to/save/unsued/images \
--id_model /path/to/arcface/model
All scripts have additional arguments that could be tuned to make the pipeline more accurate. Nevertheless, it will never be perfect. Specifically, you'd probably like to apply additional filters, depending on your data. We recommend using lenient values for the scripts above (such as those set as default) and following up with a manual filtering:
python data_preprocess/3_manual_filter.py \
--images_dir /path/to/current/filtered/images \
--trash_dir /path/to/save/unsued/images
Note that low quality images will degrade the quality of the generator!
To adapt a pretrained domain-generator to a personalized-generator, run the training command below.
python train.py \
--images_dir /path/to/dir/of/ready/training/images \
--output_dir /path/to/save/results \
--generator_path /path/to/pretrained/domain-generator \
--encoder_checkpoint /path/to/pretrained/inversion/encoder
Anchors will be inferred on-the-fly and saved under the output_dir.
To use existing saved anchors for the training images, pass --anchor_dir instead of --encoder_checkpoint. Customize hyperparameters by editing hyperparams.py file.
Training is expected to take ~50 seconds per training image on a single V100.
To generate personalized images:
python generate.py \
--generator_path /path/to/personalized/generator \
--output_path /path/to/save/results \
--anchors_path /path/to/models_anchors
First, align the test images using data_preprocess/0_align_face.py.
Our image enhancement method is based on projecting the image into the latent space of the generator, similarly to inversion. Therefore, inversion and image enhancemnet applications - inpainting and super-resolution are run using very similar code.
For inversion:
python project.py \
--images_dir /path/to/dir/of/test/images \
--output_dir /path/to/save/results \
--anchor_dir /path/to/models_anchors \
--generator_path /path/to/personalized/generator
For inpainting, add --mask_dir /path/to/dir/of/masks
For super-resolution, add --sr_factor /zero/to/one/downscaling/factor
In our experiments, we found that --beta) controls the trade-off between prior and fidelity. Depending on how "atypical" is the input image - increasing
To edit real images you first need to reconstuct them from MyStyle's latent space. The first step is to use "plain" inversion, with project.py. To improve reconstruction, you can further tune the generator on the test images (as in PTI), using train.py. Adding the second stage requires more time but it obtaines significantly better quality.
Last, with the inverted alphas and generator in hand, run the edit script:
python edit.py \
--alphas_dir /path/to/alpha/results/of/projection \
--output_dir /path/to/save/results \
--anchor_dir /path/to/models_anchors \
--generator_path /path/to/personalized/generator \
--editing_direction_path /path/to/dir/with/W/editing/directions
An example script with the entire editing pipeline is provided under editing_pipeline.sh.
Any W editing direction should work in principle. The InterFaceGAN directions used in the paper (taken from PTI) are provided in editing_directions for your convinienve. Thanks to the AnyCost GAN team - directions corresponding to all labels of CelebA can be found here.
@article{nitzan2022mystyle,
title={MyStyle: A Personalized Generative Prior},
author={Nitzan, Yotam and Aberman, Kfir and He, Qiurui and Liba, Orly and Yarom, Michal and Gandelsman, Yossi and Mosseri, Inbar and Pritch, Yael and Cohen-Or, Daniel},
journal={arXiv preprint arXiv:2203.17272},
year={2022}
}
This is not an officially supported Google product.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent