11-785 Final Project

Hair-Removal Network based on StyleGAN.

Detail implementation in Train.ipynb

Table of contents

• Data Preparation
• Boundary Training
• Male hair remove
• Initial Model Training
• Female hair remove
• Final Model Training
• Test
• Conclusion

Data Preparation

See Section 3.2.1 Data Preparation and 3.2.2 Classification,

1. sample two $w^+$ latent code set $D_0$ and $D_{noise}$.
   • take dataset in ./training_runs/dataset, select 5%-10% data from HHFQ 1024 x 1024 (self.num)
   • create files to ./training_runs/dataset/{args.dataset_name}
     • D0: without noise $D_0$
     • Dnoise: add noise $D_{noise}$
2. get hair scores through a hair classifier and gender scores through a gender classifier for both $D_0$ and $D_{noise}$.
   • wp.npy, w.npy, z.npy, hair_scores.npy and gender_scores.npy
     • hair_scores, 1: not bald, 0: bald
     • gender_score, 1: male, 0: female

Boundary Training

See 3.2.3 Separation Boundary Training

Apply the separation boundary training algorithm in InterFaceGAN to train two separation boundaries in the latent space of StyleGAN2-ada to provide general directions for male hair removal and gender transformation.

• Option one: Train male hair separation boundary by InterfaceGAN on $D_0$

  • need $D_0$ have enough bald male results.

• Option two: Use the pretrained hair separation boundary.

Male hair remove

See 3.2.4 Male Hair Removal and Training, equation (1) and (2), and optimize the latent code by minimizing the full loss function $\mathcal{L}_{dif}$ obtain $\hat{w}^{*+}_m$.

Training bald male data using hair boundary, latent_space_type = 'wp'

 create files in './training_runs/male_training'
 - mask: save hair_mask (including hat) 
 - res_wp_codes, 
 - res_img: viz_result(image after synthesis and diffusion)

Initial Model Training

See 3.2.4 Male Hair Removal and Training, composing pairs of male latent codes $w^+_m$ and $\hat{w}^+_m$ to construct dataset $H_m$ and training a fully connected network $M_m$.

1. construct dataset $H_m$
   • Create files to ./training_runs/{args.mapper_name}/data
     • train.txt
     • val.txt
     • test.txt
2. traine network $M_m$.
   • Save the best model in checkpoint.
   • Save some result in logs.

Female hair remove

See 3.2.5 Female Hair Removal and Training

Final Model Training

3.2.5 Female Hair Removal and Training, training the final HairMapper $M$ on $H$ by the same procedure in section 3.2.4.

Test

To begin real picture editing, we utilize the e4e encoder to encode the real image and obtain its latent code. The latent code is then fed to M, and the resulting latent code and image are obtained. Finally, we use Poisson editing to mix the generated image with the original face to get final result.

Conclusion

Application: Hair removal methods have two main applications. The first is digit hair design which is after success hair removal, digit hair design will be easy, as new hair template can directly apply to the result image. The second is single view face reconstruction which means hair removal provides clean textual of face without hair can use little geometry change to reconstruct 3d face view by one picture.

Limitation: Some portraits that cannot be precisely by having similar color with background. For example, in a picture, a girl with pink and purple hair which is really similar with pink and purple flower in background. Therefore some part of hair mask not be detected rather than deal with background. Shadow removal is also not be considered. The shadow caused by hair in some figure is not removal.

Future work: Improving the StyleGAN encoder or applying an additional diffusion method to improve portraits. Increase details in the supplementary should be satisfied in the additional diffusion or improved encoder. A more precise hair mask can access by a better extractor or handwork, which is helpful to improve precise. Besides, a better hair mask detected by human could also be learned to improve model performance.