A new version of artificial intelligence (AI) is available in the cartoon line of the fire! Unsupervised training, better results | code + Demo, aidemo

A new version of artificial intelligence (AI) is available in the cartoon line of the fire! Unsupervised training, better results | code + Demo, aidemo
Compiled by Xia yianne
Produced by QbitAI | public account QbitAI

Create a favorites for your favorite anime image, which will collect all of her images ...... You know, who have a few cute anime girls.

Some hand-drawn lines are cute, but black and white colors are always monotonous.

△Remember this line draft

Half a year's front-line draft color AI style2paints shell had made automatic coloring fire for a while. The good news is that the upgraded version of style2paints 2.0 was also available yesterday!

The Demo is also released. You can try it out at will ~

△Demo Interface

According to the author, style2paints 2.0 has better results than a generation, and the model training process is unsupervised.

Amazing results

You only need a line draft, and click it to give your favorite cute girl a color.

△Conversion effect of the above line draft

Awesome! Not only that, but also another color reference image can be converted into another color style.

For example, enter the following reference image:

The figure above turns into the following effect --

You can also select a color from the reference graph, and click on a certain area of the online draft with the pen, AI will refill as prompted --

There are more than one linear coloring software that supports fine-tuning, but the author said that compared with similar software, style2paints's "tip pen" is the most accurate, and users can use 3 × 3 handwriting, control the area of 13*13 on the 1024*2048 screen.

To ensure the final painting color coordination, style2paints also has a little bit of confidence: the user can only select from the reference image instead of entering the color by himself.

△Tutorial

Not simple

After you have had a good time, you may say that we have seen style migration a lot ~

However, changing a black-and-white photo into a color photo is different from converting a line draft without shadow highlight into a color picture.

This sketch composed of pure lines does not contain the brightness and texture of the pattern. That is to say, AI needs to automatically make up the information.

What's more, the authors of style2paints have high expectations for the coloring of the online draft. It's not just about entering some color between online Bars:

In a good cartoon, the eyes of sister paper should shine like a galaxy, the cheeks should be filled with a red halo, and the skin should be exquisite and charming.

△This is probably the effect.

How can this be done!

Style2paints 2.0 has not yet been published in related papers, and only the code is available for reference. However, the first version launched in June this year is actually quite effective, Achieving Color prompts Based on Semantic Information migration, making the coloring effect more harmonious.

For more information, see the Style Transfer for Anime Sketches with Enhanced Residual U-net and auw.iary Classifier GAN.
Lvmin Zhang, Yi Ji, and Xin Lin, three authors from Suzhou University, introduced how to apply the residual U-Net style of the set to grayscale images, and uses classifier generation of confrontation Network (AC-GAN) automatically for image coloring. The generation process is fast and the effect is good.

Thesis address:
Https://arxiv.org/abs/1706.03319

Different 2.0

Style2paints was very popular after its first version was launched, and many friends in the quantum field used it.

However, paintschainer, their main competing product, continued to iterate. Later, the coloring effect gradually surpassed style2paints 1.0. As a result, the authors failed to sit down and began to study new methods and updated the first version.

So what is the difference between 2.0 and 1.0?

Coincidentally, overseas netizens are also very concerned. The author replied on Reddit that, compared with the previous version, most of style2paints training is purely unsupervised or even unconditional.

That is to say, in the training process of this model, no other manual definition rules are added except for the confrontation rules, and no rules are used to force the generator's neural network to draw pictures according to the line, instead, we discovered through neural networks that, if we followed the line draft, we would be more likely to cheat the identification tool.

Similar models, such as pix2pix and CycleGAN, add l1 loss to the learning object to ensure convergence. The data received by the validator is paired with [input, training data] and [input, fake output]. The learning objectives of the style2paints 2.0 model are exactly the same as those of the classic DCGAN model. No other rules are added, and the authenticator does not receive paired output.

The author said that it is difficult to converge such a model, not to mention the depth of neural networks.

But you can see that the results are not bad.

Online draft coloring

There are actually a lot of online draft coloring procedures, such:

Paintschainer
Https://paintschainer.preferred.tech/index_en.html

Deepcolor
Https://github.com/kvfrans/deepcolor

Auto-painter
Https://arxiv.org/abs/1705.01908

In addition to paintschainer, the authors of other similar products are not very familiar with it.

He said that many Asian papers claim to be able to migrate the cartoon style, but after carefully reading the paper, they will find that their so-called "new method" is a modified VGG, although VGG generally performs well in style migration tasks, it is not very effective in cartoons.

You still have to rely on GAN, and you have to allow users to upload reference images of the style. You can't select one from Chennai Van Gogh like Prisma, and Van Gogh won't draw cartoons.

Try?

Demo:
Http://paintstransfer.com/

The thesis hasn't come out yet, but you can check the source code:
Https://github.com/lllyasviel/style2paints

The author introduces the previous version:
Https://zhuanlan.zhihu.com/p/29331219

-End-

Join community

The 13th Group of the quantum AI community has begun recruiting. You are welcome to join the group with the assistant qbitbot2;

In addition (Self-driving, CV, NLP, machine learning) Recruiting engineers and researchers in related fields.

Add the assistant number qbitbot2 to the group and note the keywords of the group ~ After the application is approved, we will invite you to the group. (Strict review by professional groups, please forgive me)

Sincere recruitment

Quantum bit is recruiting Editors/reporters working in Zhongguancun, Beijing. We look forward to talented and enthusiastic people joining us! For details, please reply to "recruitment" on the QbitAI dialog page.

Quantum bitQbitAI toutiao.com contracted

Tracking AI technologies and new product trends

Copyright Disclaimer: This article is an original article by the blogger and cannot be reproduced without the permission of the blogger.