Mengchen fish and sheep hair from the Temple Temple
Quantum position report | Public account QBITAI

Teacher Yu Hewei, is it still the feeling that you are familiar with “then play and dance”?

and if Rong Yan also entered this dimension, the style of painting even …

Ah, it feels right for the world’s high people. Rong Yan is indeed a childhood god!

Then what kind of scene will it be?

First look at Lei Jun, and the sword eyebrow star, romantic.

Look at the three big men of BAT …

Good guy, I feel that I can read the picture and write directly, and go to the orange light game “The Three Heroes of the Internet”.

(Wan -character abdominal draft has been in place …)

Crooked nuts can also be unified by this dimension.

Look at the Turing Award winner, deep learning three giants, bengio mad, Hinton is perseverance, Lecun is cold and proud, the peerless master is full, and Huashan can talk about the sword at any time.

△hinton big guy novels male lead face hammer

Even Lecun himself couldn’t help reposting:

Presumably you can see that this is indeed GAN’s masterpiece.

But this GAN rookie from byte beating is not only the ability to stand as an orange light game.

cartoon wind:

Oil painting wind:

and even Trump …

As long as 100 photos of men and women are used as training samples, letAgileGANLooking at 1 hour, it can be handy.

Even if the person in the photo wears a mask, you can make up for the face:

will also automatically convert the hat into hair. The more hairs you wear, the dense your hair is. If you wear 5 layers of hat, this is the case:

can even develop some ghost animal gameplay, such as feeding the generated images back …

△lecun becomes beautiful

and cultivated such a GAN with such a texture and martial arts ghost animals are byte beating and Nanyang University of Science and Technology. One of Song Guoxian, currently as a research intern in byte beating.

and Agilegan’s related papers have been selected as Siggraph 2021.

Only 100 pairs of sample training for 1 hour

The reason why it is namedAgileGAN(Agile GAN) is because it trains the time on a V100It only takes 1 hour, the training dataset only needs to be about100 pairs of samples(100 pieces of men and women).

So strong, how do you do it?

must know the difficulty of style migration, that is, the face like photos to cartoonsGeometric shape changes greatlyMigration.

If you emphasize the preservation of geometric shapes too much, it will cause distortion and flaws that do not meet the aesthetics.

But the reserved is not like the input photos after migration.

△The previous algorithm was either green or distorted by facial features

This is because the style migration algorithm such as Stylegan2, which encodes the features of the photo into a vector, and the inverted mappinghidden space (Latent Space)。

Based on this, the vector is changed, and then the image is maximized to generate the effect of addition and subtraction and transform gender.

△4 4 GAN INVERSION: A Survey

But the Agilegan team found that looking for the best hidden space mapping like STYLEGAN2 is incompatible, because the mapping suitable for real photos does not necessarily apply to other styles.

agilegan is improved based on stylegan2, and the solution is divided into divided intotwo parts。

The first one islevel change division self -encoder(Hierarchical Variational EutoenCoder, referred to as HVAE).

While ensuring the mapping hidden space distribution that conforms to the original Gaussian distribution, one of the original hidden spaces into multiple hidden spaces with different resolution can better encode the details of different levels in the image.

The second is starting with Stylegan2’s pre -training weight, re -tune a oneThe generator of attribute perception。

Includes multiple generating paths and multiple discriminators of different attributes (such as gender, age) to better realize the style of dependent attributes.

STYLEGAN2 generator and attribute perception of the two training stages are independent and can be trained in parallel.

This separate operation not only reduces the required training dataset size, but also makes style migration more flexible.

But the discriminator is easy to fit when using a small data set. The solution is to add an early stop strategy. Once the stylization effect meets the expectations, the training will be stopped.

This is not over yet. If you use the FIRST Order Motion, Agilegan can also complete the video migration of the video.

Byte Beat Intern One Works

In addition, Agilegan is still a “internship work”, which is formed during the period of Song Guoxian during the byte beating internship.

Song Guoxian, graduated from the University of Science and Technology of China University of Science and Technology, is currently studying a PhD in computer science at Nanyang University of Technology. At the same time, he is also an intern at the American AI Laboratory.

His research direction is mainly computer vision and computer graphics, including image -based 3D face reconstruction/analysis, VR/AR applications, and so on.

So, in Agilegan’s eyes, what is Song like?

hair volume and hair quality are really excellent.

Speaking of it, maybe you can play such a GAN directly on Douyin.

If you can’t wait, the author has released a trial version:

http://www.agilegan.com/

Thesis address:
https://guoxiansong.github.io/homepage/paper/AgileGAN.pdf

Project address:
https://guoxiansong.github.io/homepage/agilegan_cn.html

Reference materials:
[1]https://www.researchgate.net/publication/348487325_GAN_Inversion_A_Survey
[2]https://mp.weixin.qq.com/s/ayt6g-5KoSV14s6a5mp9pg