https://github.com/pd-mera/anime-face-generation-stylegan2-ada
Generating Anime Faces with StyleGAN2-ADA
https://github.com/pd-mera/anime-face-generation-stylegan2-ada
Last synced: 3 months ago
JSON representation
Generating Anime Faces with StyleGAN2-ADA
- Host: GitHub
- URL: https://github.com/pd-mera/anime-face-generation-stylegan2-ada
- Owner: PD-Mera
- Created: 2023-01-17T03:00:30.000Z (over 2 years ago)
- Default Branch: main
- Last Pushed: 2023-03-12T07:54:46.000Z (about 2 years ago)
- Last Synced: 2025-01-08T19:46:58.685Z (5 months ago)
- Language: Python
- Size: 14.3 MB
- Stars: 2
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# Anime Face Genaration with StyleGAN2-ADA
Generate Anime Face with StyleGAN2-ADA
## Environments and Dependencies
- Python 3.8.16
Install requirements
``` bash
pip install -r requirements.txt -f https://download.pytorch.org/whl/torch_stable.html
```
## Data
Put raw images in `./data_raw/` and run
```
bash prepare_data.sh
```
## Train
If you want to train from scratch, modify hyperparameter in `train.sh` and run
```
bash train.sh
```
Or if you want to continue training from checkpoint, modify hyperparameter in `train_resume.sh`, especially `RESUME_NET`
* **Set Initial Augmentation Strength**: use `--initstrength={float value}` to set the initialized strength of augmentations (really helpful when restarting training)
* **Set Initial Kimg count**: use `--nkimg={int value}` to set the initial kimg count (helpful with restarts)
```
bash train_resume.sh
```
For more training config, refer to [training](https://github.com/NVlabs/stylegan2-ada-pytorch#training-new-networks) in stylegan2-ada repo
- If you want to visualize training losses and scores, you can modify training-runs in `visualize.py` run
```
python visualize.py
```
You can get a `results.csv` in `./csv_results/` folder. Visualize that data by your own
## Generate Image
You can download pretrained weights from [here](https://drive.google.com/file/d/1y12u6_3ORAr4Gn-F-jovtONU1deLBWOV/view?usp=share_link)
### Generate Single or Multiple Images
Modify hyperparameter in `generate.sh` and run
```
bash generate.sh
```
### Generate Style Mixing Images
Modify hyperparameter in `style_mixing.sh` and run
```
bash style_mixing.sh
```
### Project Image to Latent Space and Reconstruct
- Project Image to Latent Space
```
bash latent_projector.sh
```
- Generate Image from Latent Space
```
bash latent_generator.sh
```
| Root Image | Generate From Latent |
|---|---|
|  |  |
**Project Process**
You can find projecting process in better quality in video [here](https://drive.google.com/file/d/1YyMm_zPxt8eHxrHa-i7Ka0u9nXweER8d/view?usp=share_link)
## Closed Form Factorization
Thanks to [dvschultz/stylegan2-ada-pytorch](https://github.com/dvschultz/stylegan2-ada-pytorch) and [pbizimis/stylegan2-ada-pytorch](https://github.com/pbizimis/stylegan2-ada-pytorch)'s for convert Close Form Factorization from [rosinality/stylegan2-pytorch](https://github.com/rosinality/stylegan2-pytorch)'s repo
### Get factor
```
python closed_form_factorization.py --ckpt [PKL_CHECKPOINT]
```
You will get `factor.pt`
### Apply factor
```
python apply_factor.py -i 50 -d 5 --trunc 0.7 --seeds r1 --ckpt [PKL_CHECKPOINT] factor.pt
```
You will get image with your own factor modification
## References
- [NVlabs/stylegan2-ada-pytorch](https://github.com/NVlabs/stylegan2-ada-pytorch)
- [rosinality/stylegan2-pytorch](https://github.com/rosinality/stylegan2-pytorch)
- [dvschultz/stylegan2-ada-pytorch](https://github.com/dvschultz/stylegan2-ada-pytorch)
- [pbizimis/stylegan2-ada-pytorch](https://github.com/pbizimis/stylegan2-ada-pytorch)