Zhejiang big sister made a bald generator to board CVPR, perfectly retaining the facial features of the face, which adds extra points to the resume

Xiao Zhen is from The Temple of Oufei

Qubits | Official account QbitAI

Still shaving your bald head at a barbershop to find a job? Try this generator!

Just enter a photo, the output is a flawless bald head, and the temperament is immediately promoted from internship to supervisor (manual dog head)

Whether you are a young programmer brother or a programmer with long hair, you can "get stronger" through this method:

With this "bald artifact" called HairMapper, P hair is a matter of minutes, but the facial features are completely unaffected:

Well, after having a hairstyle you want to do, directly P one...

To achieve such a good effect, it is necessary to use three GANs to work together.

A "bald artifact" that everyone can use

Specifically, the principle of HairMapper is divided into three steps, generating a bald head that retains five features and merges avatars.

First, the authors used StyleGAN to create a baldness effect similar to the original head shape:

At this time, we see that StyleGAN cannot accurately control the generation of facial features, so the face shape has changed slightly.

So the next step is to use InterFaceGAN to cut out a face shape and facial features other than hair, while also preserving the surrounding scenery:

Finally, the effect generated by the first two steps is synthesized to make a perfect "bald head":

However, it is worth mentioning that due to long hair, small data sets of female bald heads, etc., the authors added an additional step to the female version of the "bald head artifact".

Among them, the female photo must first be converted into a male photo through StyleFlow, and then repeat the above steps to finally get a bald head:

This effect is still good, and tests in the dataset show that even girls with amazing hair volume can lose their hair (doge) with one click.

The authors compare HairMapper to other bald generators.

It also made a dataset of 6000+ bald heads

In fact, it's not that no one has done a baldness generator before, it's just that the effect is not ideal.

For example, this generator called StyleFlow that Adobe has previously launched can make Musk bald:

But one is that the hair is not clean, and the other is that the facial features have also given GAN a "nother person". In addition to StyleFlow, other types of "bald generators" also face this problem:

This time, HairMapper retains the original facial features on the basis of GAN hair loss, making the restored face very real.

Whether men and women young or old or round flat pointed toe:

Or a variety of different skin tones and even different skin types:

Even the side face and even the shadows can get the silky bald head effect under the action of HairMapper:

The authors have used this algorithm to turn 6,000 portraits from the Flickr-Faces-HQ (FFHQ) dataset into bald heads.

In the future, whether it is P hair or transforming into a men's/women's clothing gangster, you can use this open source dataset to train:

About the author

Yiqian Wu, from zhejiang university's computer-aided design and graphics (CAD&CG) laboratory, is currently a phD student with interests in computer vision and portrait editing.

Ersaku Yong-Liang Yang, assistant professor at the University of Bath, graduated from Tsinghua University with a ph.D. and studied under Professor Hu Shimin, whose research interests are computer graphics, geometric modeling, VR/AR, etc.

Corresponding author Jin Xiaogang, professor of CAD &CG Laboratory of Zhejiang University, whose research interests include film special effects simulation, automatic driving simulation, and creative modeling.

Can't wait to try the "bald artifact"?

According to the authors, HairMapper's pre-training code is still in the process of being expedited and will be released soon, and the dataset can already be used