AI graphic art the magic behind the Prisma method Depth

Lei feng’s network: the writer is using technology engineers.

Prisma applications download the comments section to see such a funny Word: half the world’s people use the catch the spirit of the PokemonGO and the other half were using this software.

Image Edit software early countless, like film wind of VSCO, puzzle beauty of took State have, and beauty tuxiuxiu, two times Yuan cartoon of magic diffuse camera, and face MoE, self-timer beauty of PICSPLAY, and beauty camera, main filter camera of Roookie Cam, and MIX filter master and so on, these phenomenon level products in this fast Diego generation of times endlessly, so Prisma by what in such of environment Xia deus ex? Even Russia’s Prime Minister Dmitry Medvedev make a “flop”.


AI graphic art: the magic behind the Prisma method | Depth

First, Prisma not simple to in original Shang conversion tones or environment style, also not just using General of filter to change image of light, and color or overlap pictures, it will according to you provides of pictures content and you requirements of style pictures of style again “painting out” a Zhang new of pictures, it does not must makes you variable beauty, but it must let you met new of himself (following, left a for original).

AI graphic art: the magic behind the Prisma method | Depth

So, Prisma and algorithms of artificial intelligence Neural art what’s going on? Prisma and CNN combined, generating pictures what is the process?

| Prisma Neural and artificial intelligence algorithm art

Prisma style conversion is done by machine vision research in the field has been a problem, called a texture transfer (texture transformation), target-style synthesis of the source image to the target while retaining the semantic content of the target picture in picture.

In essence, the picture content and style separation is not a well-defined question (Well-defined problem), this is because does not define exactly which part of the picture, which part belongs to the style, content and style is most likely intertwined, there is no way to separate.

So the content and style can be completely separated it is still an unanswered question, if you can, then what kind of approach?


AI graphic art: the magic behind the Prisma method | Depth

The technology behind the Prisma

Behind the Prisma technology is based on a year CVPR oral article “Image Style Transfer Using Convolutional Neural Networks”.

Actually detonated Prisma mass market six months ago, the ingenious and amazing results of this method has been in the academic world has aroused wide concern. As is well known in the field of artificial intelligence now causing deep learning technology revolution has many tasks to a practical level, and Prisma technique is also not surprisingly based on neural network and its core system is characterized by nerve to separate, and combine random pictures of content and style, in order to implement an algorithm that can be used to describe art images. Its basic idea is to use a multi-layer Convolutional neural networks (CNN) abstract a given painting some hidden features that advanced to imitate painting styles and applied to this style of painting a new picture. Just Cavalli iPhone 5 Case

Just Cavalli iPhone 5 Case

Texture transformation areas, traditional methods have failed to achieve such amazing results, its essence is not the deep learning method can only be gets to the target before pictures of image low-level features, which led to these methods cannot be independent of the meaning of the picture content and effective modeling style change, which cannot be good for both decoupling and synthesis.

Deep neural network can achieve amazing results in many areas, just because it can extract a high level of information representation. Prisma methods can succeed, is a clever use of deep expression of neural network taking senior pictures, in a matter of 10 seconds to turn an ordinary photo into a pair of highly artistic features of modern paintings or works of Western painting styles.

Where is the secret?

The following sketch of two of the papers will be used in the complete analysis of the secret.

As said above, the style transformation needed to separate the content and style of the picture that, shown respectively in the source image (style) and objectives (content provider) to consist of layers of convolution and pooling layer depth in neural networks. For target directly using convolution the response at each level in the reconstruction, visualization of results results in the red box, can be seen in the lower image and target figure almost unanimous, and high level network reconstruction of some details of the image pixels are discarded and those picture-level semantic content is retained. Source image convolution of the characteristics of each layer (feature maps) the correlation coefficient to reconstruct the style feature, Visual results can be seen from the green box, this pump style expressed in different network layer successfully extracted and scale characteristics.

AI graphic art: the magic behind the Prisma method | Depth

Figure explains how deep neural network on style and content separate models, then you can use the depth of supervised learning neural network-style converted.

Figure in the around on both sides of network for extraction source figure of style said and target figure of content said, and middle of network for on style for synthesis, papers using of is white noise pictures as Kai began figure, thought is through around two a network provides of style and content characterization for supervision learning, makes entered pictures each layer extraction out of style characterization and senior extraction out of content characterization and around two a network corresponding network layer reconstruction out of characterization increasingly consistent, so through standard of random gradient declined algorithm, Iterative synthesis makes white noise pictures eventually want.

Using mathematics expression to description above thought is need structure a loss function (red box by shows), this loss function is by style loss items and content loss items both linear composition, which alpha and BelTA respectively representative style and content of weight proportion, if Alpha/BelTA high is generated figure will more highlights out content and style of will less some, low is generated figure will style of strongly but content was diluted, this is Prisma provides user regulation of principle where. (Loss of style and content loss of details you can refer to the paper. ) When there is a loss of function, can be solved for image vector gradient, then use the bottom-transfer algorithm can update the input image continued to style conversion (green box).

In addition, by extraction of image features changes before the size of the image, can be controlled by-style_scale parameter which art features extracted from the image. (Three images from left to right,-style_scale=2.0, 1.0, 0.5)

AI graphic art: the magic behind the Prisma method | Depth

In addition, the Prisma can also use more than one style images to blend a variety of artistic styles. (Below are as follows: “The Starry Night” + “The Scream” and “The Scream” + “Composition VII”)

AI graphic art: the magic behind the Prisma method | Depth

Or use more than one style when the image, you can control the degree of mixing different types of images.

AI graphic art: the magic behind the Prisma method | Depth

Also can change color under the premise of image styles, for example, if you set-original_colors 1, system retains the original color of the picture (below).

AI graphic art: the magic behind the Prisma method | Depth

Above is all behind the Prisma technology principle, when we analyze discovered that in fact the principle is not complicated, for people familiar with deep learning, in less than a week will be able to replicate the algorithm.

But the most important thing is to realize this algorithm is the key to success is to use the ability to express his deep learning network high-level semantic information, and cleverly constructed loss function, just make use of other deep neural network for solving it.

| Using AI algorithm to graphic was cool, but Prisma also has some drawbacks

Like, must to in has network state better, and network more fast of State Xia only can using smooth, because need online loaded painting style image and user provides of content pictures need sent to remote of server Shang for artificial intelligence calculation Hou again biography back, so, relative other repair figure software, it of pictures edit speed slow many, General are need waiting for several 10 seconds only can see preview effect.

AI graphic art: the magic behind the Prisma method | Depth

In addition, because all image processing is done in the cloud, so there will be a small group of users to use an excessive number of server overload.

AI graphic art: the magic behind the Prisma method | Depth

| Use of deep learning, why don’t deepart.IO fire like a Prisma?

Before the advent of the Prisma, also had a graphic application based on advanced learning technology: deepart.IO, but deepart cannot achieve success like a Prisma.

 One is the GPU of a picture takes a lot of time, leads to a lot of server resources, if the user is willing to bear the cost, the processing time in half an hour or so, if not donors, needed about 6 hours;

The other hand, the researcher at the University of deepart is mainly made of technical experiments, and will not invest too much energy, in the style of the image type, brand promotion, and commercial applications is considered little.


AI graphic art: the magic behind the Prisma method | Depth

Prisma by optimizing algorithms, most of the time is relatively short, and users do not need to the page to submit photos, do not need to register to use free filter 35 different painting styles, in this era of mobile Internet, Mobile App is more popular.

Recently, the Prisma launched an Android version, which will no doubt bring more users, daily activity could surpass 2 million people. According to reports, the “video filters” has entered beta, released about a week later and, in addition, may also be introduced “art filter GIF”, “video” and so on, creating a “flow of oil painting.”

Remember two years ago, Poland painter-Director Dorota kobiela, and Oscar-winning producer Hugh welchman began preparations for a special tribute to Xiang Fangao movie loving hand-painted oil painting of Vincent, hundreds of artist after high intensity training and daily copy of Van Gogh’s style, has completed more than 56,000 paintings. In this film, every static picture is imitating van Gogh drawing style of painting, film every minute of the day 12 this picture will be used, and then a sequence of still pictures together in quick play, form a dynamic effect.

If art Neural technology successfully applied to video, this van Gogh’s film did not need such high investment, and even “shoot” any style of movie, movie and art blend and seal himself in his art, is a very beautiful thing, like the Mona Lisa, Leonardo da Vinci’s painting, her smile is eternal. People’s pursuit for beauty and timeless, this maybe the Prisma can be quickly captured the hearts of reason. Just Cavalli iPhone 5 Case

Prisma’s significance lies in the name of art with a popular way to show deep learning capabilities to the masses. While many companies use AI techniques of artificial intelligence to serve the public, but so hot like Prisma set off public is surprised now artificial intelligence technology is also very rare.

But I think this is just a start, AI technology is in every area, both high-tech fields such as autopilot, robot, or some areas, such as art, social welfare will continue to see benefits from artificial intelligence and surprise.

Maybe when you first use the Prisma was deeply impressed by the effect. I want to say is, your surprise has just begun.

Lei feng’s network (search for “Lei feng’s network”, public interest) Note: drawings from the This is Lei Feng a net is exclusive to writing articles, please contact licensing and label source and author, no deletion of content.


Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google+ photo

You are commenting using your Google+ account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s