AI helps human painters win first place in art competitions. What's the secret behind it?

Two years ago, in the Colorado State Fair art competition, a work called "Space Opera" won first place.

This painting is majestic, light and dark, and is quite reminiscent of the French symbolist painter Gustave Moreau. However, it was not drawn by a human, but by a contestant with no drawing foundation, using an AI drawing tool.

Let’s turn the time back to 2018. At that time, an AI painting called "Portrait of Edmund Bellamy" sold for more than $400,000 at Christie's auction house in New York. This is the first artificial intelligence work to be auctioned, which also marks the beginning of AI art works being recognized by the market.

Nowadays, AI painting has become commonplace, and AI players at home and abroad are leaving the field one after another, leaving this track in smoke.

For example, the well-known foreign AI photo-generating applications Midjourney, Stable Diffusion, and DALL-E are constantly fighting with gods; domestic companies such as Byte, Alibaba, and Tencent also launch new products and new functions every now and then.

However, there are various AI painting applications. For AI enthusiasts, how to achieve stronger image-text matching? How to generate more aesthetically pleasing images? How to efficiently deploy the Vincent diagram model?

The sixth episode of "AIGC Experience School" will talk about this matter.

The "AIGC Experience Party" video program is jointly launched by Volcano Engine, NVIDIA, this website and CMO CLUB, bringing together many professionals in the marketing field to discuss the application, thinking and practice of AIGC technology.

In this episode, two industry senior guests will deeply analyze the technical links behind high-esthetic text images and teach everyone how to achieve stronger image-text matching.

• Live broadcast time: August 6th 19:00-19:40

• Live broadcast theme: Fantasy imaging, the technical link behind high-beauty Vincentian pictures

Sixth This episode of the program invited two important guests, namely Doubao Wenshengtu technology expert Zhang Liang and NVIDIA solution architect Zhao Yijia.

Preview of exciting content

With the iteration of AIGC technology, the multi-modal capabilities of generative AI are rapidly expanding and deepening. As one of the hottest application areas for large models, Vincentian graphics have also made breakthrough progress in recent years, greatly enriching the arsenal of content creators.

Users only need to enter a few simple prompt words, and AI can automatically generate realistic and delicate paintings. Whether it is the delicate texture of classical oil painting, the free expression of modern abstract art, or the profound artistic conception of oriental ink painting, it can create with ease. It can not only learn and imitate the styles of famous painters in history, but also create new art genres, bringing users unprecedented creative and interactive experiences.

Today, AI painting has evolved to become "sophisticated", even to the point where it is difficult to distinguish between true and false.

For example, this:

or this:

It is precisely because of its unique creativity and immersive interactive experience that AI Wenshengtu has become the "fragrance" that major AI manufacturers are competing for. cake".

For example, Byte has launched AI products such as Jiuying, Xinghua, and Jimeng in one fell swoop. Without exception, they all include Wenshengtu capabilities, and behind these products are inseparable from Byte’s Doubao Wenshengtu Large model.

For the Vincentian graph model, there are nothing more than three things to win the "popularity" of users: stronger image-text matching, more beautiful image effects, and a faster generation experience.

So, how does the Doubao Wensheng large model achieve stronger image and text matching to meet the user’s idea design? How to generate more aesthetic images to provide a more ultimate user experience? How to produce graphs more quickly to meet extremely large-scale service calls?

The answers to these questions will be revealed in the sixth issue of "AIGC Experience Party".

Live broadcast registration

At 7pm on August 6th, come to the live broadcast room to watch "Fantasy Imaging, the technical link behind the high-beauty Vincentian pictures", scan the QR code to go directly to the live broadcast room.

Review of the last issue

In the fifth live broadcast of "AIGC Experience Party", Zhang Jianyang, head of content security and risk control of Volcano Engine, and He Kun, senior manager of NVIDIA enterprise-level developer community, gave an in-depth analysis of AIGC in the marketing field innovation and security issues, and share how to achieve sustained growth and steady development of business amid changes.

As we all know, AIGC technology is gradually becoming an innovative tool for corporate marketing. It can help marketers create imaginative marketing materials at an unprecedented speed. By simply entering a few keywords or prompt words, attractive marketing copy, pictures, and marketing videos can be automatically generated in a short time.

However, just like any technological innovation, while AIGC improves marketing creativity and efficiency, it also brings many security issues.

Zhang Jianyang, head of content security and risk control at Volcano Engine, used actual cases to demonstrate how Volcano Engine’s AI technology can “protect” marketers.

Zhang Jianyang said that the illusion of large models makes the output results not completely controllable, which leads to various potential content risks in marketing content. For example, the marketing copy created by AIGC technology contains "absolute" terms such as "best", "first", "world-class", "first on the entire network", etc., and the generated marketing images or videos may contain vulgar and low-quality words. Bad content, once it appears in advertisements, will not only damage the brand image, but also cause public disgust.

In this regard, Volcano Engine provides a complete set of targeted content risk detection engines. Through its multi-modal content support, diversified risk models and real-time risk assessment functions, it creates a comprehensive and efficient risk management for enterprises. solution.

Among them, to address the issue of risk diversity in marketing materials, the detection engine has built-in hundreds of risk models. For example, the advertising terminology detection model suitable for marketing copy can accurately determine whether the copy contains absolute terms, false propaganda, and illegal content that induces consumption; while the picture quality and copyright detection model suitable for marketing videos can determine whether the copy contains absolute terms, false propaganda, and illegal content that induces consumption. Whether it contains bad images and low-quality content, and whether there is any possibility of infringement.

In addition, in order to effectively convert target users, companies often invest heavily in various marketing activities, such as newcomer rewards, task rewards, etc., in order to increase user stickiness. However, these tempting reward mechanisms may also become "fat meat" in the eyes of illegal companies. They use loopholes to cheat and take the company's marketing funds for themselves.

For example, when a car company carried out online marketing activities, it encountered two thorny problems: false assistance and false test drives. After accessing the Volcano Engine intelligent risk control product, in just two months, they effectively identified 1.08 million abnormal behaviors, more than 30,000 abnormal accounts, and more than 20,000 abnormal devices, and the direct stop loss amount reached 500,000 yuan. Not only does it effectively reduce the waste of marketing expenses, but it also greatly improves the business conversion rate.

He Kun, senior manager of NVIDIA's enterprise developer community, introduced a cool tool - NeMo Guardrails, which is like a "bodyguard" for large language models to prevent them from talking nonsense.

According to He Kun, NeMo Guardrails is an open source tool library that controls the output of large language models by adding programmable guardrails to prevent sensitive topics, inappropriate remarks, and malicious attacks. It includes three major guardrails: topic limitation, conversation security, and attack defense. It supports custom conversation logic and ensures that the response of the conversation robot complies with specific requests and predefined logic.

The workflow of NeMo Guardrails includes receiving user input, matching generated guardrails, executing conversational logic flow, and generating bot answers. It defines conversational logic using Colang, a language for conversational AI modeling that is highly readable and scalable. In addition, NeMo Guardrails can be used in conjunction with LangChain to support multiple large language model dialogue engines to achieve more secure dialogue system output.

The fifth issue review viewing address: https://vtizr.xetlk.com/s/7CjTy (or identify the QR code below).

Reward Questionnaire

Have you currently used AIGC in your daily work? We welcome your feedback. We have prepared some small gifts, please remember to fill in the mailing information.

Click to read the original text to participate in the questionnaire.

The above is the detailed content of AI helps human painters win first place in art competitions. What's the secret behind it?. For more information, please follow other related articles on the PHP Chinese website!