The task of StyleGAN V2 is image generation while the Editing module uses the attribute manipulation vector obtained by pre-classifying and regressing the style vector of the multi-image to manipulate the attributes of the generated image.
## How to use
### Editing
The user can use the following command to edit images：
```
cd applications/
python -u tools/styleganv2editing.py \
--latent <PATH TO STYLE VECTOR> \
--output_path <DIRECTORY TO STORE OUTPUT IMAGE> \
--weight_path <YOUR PRETRAINED MODEL PATH> \
--model_type ffhq-config-f \
--size 1024 \
--style_dim 512 \
--n_mlp 8 \
--channel_multiplier 2 \
--direction_path <PATH TO STORE ATTRIBUTE DIRECTIONS> \
--direction_name <ATTRIBUTE NAME TO BE MANIPULATED> \
--direction_offset 0.0 \
--cpu
```
**params:**
- latent: The path of the style vector which represents an image. Come from `dst.npy` generated by Pixel2Style2Pixel or `dst.fitting.npy` generated by StyleGANv2 Fitting module
- output_path: the directory where the generated images are stored
- weight_path: pretrained model path
- model_type: inner model type in PaddleGAN. If you use an existing model type, `weight_path` will have no effect.
Currently recommended use: `ffhq-config-f`
- size: model parameters, output image resolution
- style_dim: model parameters, dimensions of style z
- n_mlp: model parameters, the number of multi-layer perception layers for style z
- channel_multiplier: model parameters, channel product, affect model size and the quality of generated pictures
- direction_path: The path of the file storing a series of attribute names and object attribute vectors. The default is empty, that is, the file that comes with ppgan is used. If you don’t use it, please remove it from the command
- direction_name: Attribute to be manipulated，For `ffhq-conf-f`, we have: age, eyes_open, eye_distance, eye_eyebrow_distance, eye_ratio, gender, lip_ratio, mouth_open, mouth_ratio, nose_mouth_distance, nose_ratio, nose_tip, pitch, roll, smile, yaw
- direction_offset: Offset strength of the attribute
- cpu: whether to use cpu inference, if not, please remove it from the command
- latent2: The path of the second style vector. The source is the same as the first style vector
- weights: The two style vectors are mixed in different proportions at different levels. For a resolution of 1024, there are 18 levels. For a resolution of 512, there are 16 levels, and so on.
The more in front, the more it affects the whole of the mixed image. The more behind, the more it affects the details of the mixed image. In the figure below we show the fusion results of different weights for reference.
- need_align: whether to crop the image to an image that can be recognized by the model. For an image that has been cropped, such as the `src.png` that is pre-generated when Pixel2Style2Pixel is used to generate the style vector, the need_align parameter may not be filled in
- start_lr: learning rate at the begin of training
- final_lr: learning rate at the end of training
- latent_level: The style vector level involved in fitting is 0~17 at 1024 resolution, 0~15 at 512 resolution, and so on. The lower the level, the more biased toward the overall style change. The higher the level, the more biased toward the detail style change
- step: the number of steps required to fit the image, the larger the number of steps, the longer it takes and the better the effect
- mse_weight: weight of MSE loss
- pre_latent: The pre-made style vector files are saved to facilitate better fitting. The default is empty, you can fill in the file path of `dst.npy` generated by Pixel2Style2Pixel
- output_path: the directory where the generated images are stored
- weight_path: pretrained model path
- model_type: inner model type in PaddleGAN. If you use an existing model type, `weight_path` will have no effect.
