Your Double eleven. Buy, buy, buy.

Ali poster designer’s double 11, in a named “resource bit group” small dark room overnight overtime.

Make posters, change text, change products, adjust design, change banner, each designer docking with several operators, foxconn assembly line like repetitive work. On November 11, we completed hundreds of millions of posters.

However, this is becoming a thing of the past…

AI changed Go, and now it’s changing poster design.

This is an AI designer named lu Ban, and yes, he will be responsible for the design of 400 million banner posters for This year’s Double 11.

But considering lu Ban can design 8,000 posters in one second on average, or 40 million in a day, 400 million is only a small goal.



ruban

The poster design AI “Lu Ban” was also born in connection with Double 11.

Every year, Double 11 is a big test for Ali designers: the massive design needs, the need to ensure that all the people are unified and standardized, and the urgent design needs, and the rapid draft, are both mental and physical tests for every designer.



Therefore, after November 11, 2015, Ali began to have ideas inside. In that year, Ali officially implemented “thousands of faces” in product recommendation, so it is hoped that the design of strong marketing oriented advertising resources can also achieve “thousands of faces”.

Therefore, “Luban” project was formally established and developed to the present “Ali Intelligent Design Laboratory”.

At a time when AlphaGo was sweeping the world and spreading deep learning and AI, Alibaba internally decided to further build Lu Ban into an ALPHAGo-like AI designer.

And then began to build a neural network, so that Lu Ban learned from the achievements and experience of human designers, and continued to evolve. By November 11 this year, Lu Ban’s level has reached the level of ALI internal P6.

Lu Ban’s learning evolution mainly has three technical principles.



Three core modules

From 0 to P6, Lu Ban’s self-taught design ability mainly depends on three modules: style learning (planning + elements), action device, and evaluation network.


The first is the style learning module.


Lu Ban first annotates the design data of a large number of design materials in a structured way, and finally outputs the spatial + visual design framework through a series of neural network learning.

In the framework design, first of all, through manual annotation, let the machine understand the elements of the design, such as its commodity body, flower background, mask.

At the next level, you also need to define some design techniques and styles through the experience of design. Gimmick refers to why these elements can be constructed the way they are.

The top layer is the style, when these elements are formed, what it looks like aesthetically or visually, so that the machine knows what it is made of.

The next step is to prepare the original documentation of the design, such as a list of flowers and design methods, for input into the deep learning network. The network has some memory function, which can remember the complicated process in the design step.

After this layer of neural network learning, will get a design framework. Technically, it’s a model of a bunch of spatial features and visual features. From a designer’s point of view, it corresponds to the general frame impression in the designer’s mind before making a set of designs.



While designing the framework, element center also inputs elements in batches (such as base map, main product map, modification elements, etc.), which are learned by element classifier and classified according to visual features and types.

Specifically, The Luban team collects some copyright libraries in advance, as well as the way they create design elements, and inputs them into the element classifier. The classifier will distribute these elements into various types, such as background, body, and decoration, and will also complete the extraction of the image library.


Then there are the action devices.


The main function of the action device is to select the design prototype from the style learning module according to the needs, and select elements from the element center, plan a number of optimal generation paths, and complete the picture design.

This is very similar to the actual working process of designers. If designers want to design a flower, they will constantly adjust every position, every pixel and every Angle in the software. At the same time, the whole process is a reinforcement learning process, the action device will become smarter and smarter through trial and error.

When this process is complete, multiple designs are output and ultimately sent to the Evaluation Network to score the output product.


Finally, the evaluation network.


Evaluation networks work by feeding in large numbers of design images and rating data, and then training machines to judge good designs.

The foundation of Luban is derived from the designer’s design template materials and element materials, so there are two designers to train Luban every day, one is responsible for helping Luban complete the latest style learning (style learning), so that Luban continues to evolve and master better design skills.

The other role is to evaluate the results of Lu Ban’s design (evaluation network) and tell Lu Ban what kind of design is the best.

The core job of a designer is to turn design into data. At present, Lu Ban has learned millions of design drafts, has evolved into hundreds of millions of poster design ability.

In fact, as you can see from the initial design of AlphaGo, Luban goes from 0 to P6, which is also a collaborative effort of designers and algorithm engineers.

Behind this, Alibaba’s designers and algorithm engineers did three big things.


The three lessons

First, domain research. Find experts in the field to delve into their empirical knowledge and build a set of data models that machines can learn from. Visual design experts abstract design problems into a set of “style-gimmy-template-element” data models, turning years of visual design experience into machine-learnable “data”.

Second, data links. After the data model is defined, the data is captured and annotated, and the data set is classified and managed. In this process, how to process data to update the frequency of algorithm training, what data to use to verify the model, how to evaluate the effect of the model, and how to get through the offline model and online data at the product end? This series of data problems require a clear data link design.

Third, algorithm framework. Algorithms are framed by algorithmic scientists, and data and algorithms are like gasoline and engines. The product designer needs to discuss with the algorithm and input business scenarios and data problems into the algorithm.

This is also the reason why Ali internally asks product designers to learn machine learning, because only by understanding the algorithm framework and technical principles can we better understand how things work.

But it is not without specific challenges.

In the process of luban’s construction, it encountered three technical challenges.


Technical challenges

The first is the lack of annotation data. All artificial intelligence today is based on massively structured annotated data, and the design process has not even been completed online, let alone standardized, structured data.

The second is the uncertainty of design. Design is a very uncertain thing, design requirements grasp and result evaluation are human subjective consciousness. For example, you can’t type “high-end poster” into the machine.

Finally, there is no precedent. There are no off-the-shelf technologies or frameworks in the industry to refer to, unlike the benefits of AlphaGo.

At that time, after AlphaGo team published the paper, go AI all over the world improved their fighting ability according to this, such as Tencent Juyi, soon reached the world level.

But for Lu Ban, there was no previous experience to refer to. It was all on his own. However, it was not all fruitless. During the year of exploration, ali Luban team had a clearer definition of AI products.

They thought internally that Lu Ban’s AI was controlled visual generation. Controllable means to control intelligently according to the needs of business and business. Visual generation shows that Lu Ban solved the problem of visual creation from scratch.


The wheel test

So the poster design AI Lu Ban, how about the effect?

On November 11, 2016, Lu Ban made his debut. It ended up producing 170 million banner ads, increasing click-through rates by 100%.

Compare that to a human, assuming that it takes 20 minutes for a human designer to figure out each map, and it would take 100 designers to figure out each map for 300 years.

…

You don’t have to count the cost savings.

This year, Luban has been further iterated. The design level has been significantly improved, and the latest data are as follows:

Lu Ban has learned millions of design drafts and has evolved hundreds of millions of poster designs.

On November 11 this year, Luban has been able to produce 40 million posters a day and design 8,000 posters per second on average. Each poster will be specially designed according to the characteristics of commodity images. In other words, no poster designed by Luban will be exactly the same.


The Future of designers

Surely it’s time to talk about the future of the designer community.

According to the current evaluation system of technical posts in Ali, the poster design AI Lu Ban has reached the LEVEL of P6, and the subsequent progress will only be faster and faster.

Will designers be replaced by AI?

Yes, in Ali’s system, designers around P4 are “threatened” by machines.

But not all, in addition to the “creative” part of the machine helpless, human designers and machine competition, will also produce a “trainer” like the new profession.

These trainers are the core personnel of Luban’s data center. They need to provide larger-scale and richer data for luban’s evolution and realize the transformation of “structured data” for many style-related things.

Ali Intelligent Design laboratory told qubits, ali designers now, become to learn luban system, learn how to train machines, at the same time do aesthetic control.

“It took lu Ban half a month to learn the double 11 design style this year, and he has already started to produce some designs that no one has taught him. But the most innovative creative designs can only be done through human-machine collaboration.” Luban director Le Sheng said.



So, human designer friends, are you ready to spend time with master carpenter, no, AI designer Lu Ban?


via QbitAI