Our work is divided into four steps:the first step is model training;the second step is latent space traversal;the third step is attribute traversal; the fourth step is editing direction semantic identification and quantitative evaluation of disentanglement.

1. Train model for five methods

Train single discovery dirextion(SDD)

please run "train_Latent.py" file.Run the command in the "train_latent" section of the launch.json file.

Train single Hessian Penalty(SHP)

please run "hessian_penalty_stylegan.py" file.Run the command in the "hessian_penalty" section of the launch.json file.

Train Hybrid Discovery and Penalty (HDP)

please run "train_Mix_Latent.py" file.Run the command in the "train_Mix" section of the launch.json file.

Train Discovery To OroJaR(DTJ)

please run "line_2_orojarpenalty.py" file.Run the command in the "line_2_orojar" section of the launch.json file.

Train Discovery to Hessian Penalty(DTH)

please run "line_2_hessianpenalty.py" file.Run the command in the "line_2_hessian" section of the launch.json file.

2.Traversing the latent space

please run "traverse_line_latent.py",The relevant configuration parameters are in section "traverse_line_latent" of file launch.json.

3.Traversing the attribute space

please run "traverse_attribute_space.py",The relevant configuration parameters are in section "tra_attr" of file launch.json.

4.Editing direction semantic identification and quantitative evaluation of disentanglement.

please run "rank_interpretable_paths.py",The relevant configuration parameters are in section "Inter_path" of file launch.json.

5.Result and GIF

• Method SDD edits result of au_12_Lip_Corner_Puller property in direction 162
• Method SHP edits result of au_12_Lip_Corner_Puller property in direction 98
• Method HDP edits result of au_12_Lip_Corner_Puller property in direction 170
• Method DTJ edits result of au_12_Lip_Corner_Puller property in direction 6
• Method DTH edits result of au_12_Lip_Corner_Puller property in direction 86
• Comparison of SDD and HDP accuracy

Related PDF Papers

The relevant papers have not been formally published and are not readily available to the public.