After the blessing of large models, are digital people 'more human'?

Beijing Winter Olympics AI virtual human sign language anchor, Hangzhou Asian Games digital human ignition, Xinhua News Agency digital reporter, digital astronaut Xiaowei... As more and more digital humans appear in people In life, the entire digital human industry is also developing towards diversified and wide-ranging applications, rapidly expanding into different industries and different scenarios.

For the C-side, digital people help users produce content and assist with work, such as: digital people practice spoken language, play games with digital people, etc.; for the B-side, digital people It is the "tool man" of the enterprise and is used in finance, film and television, e-commerce, live broadcast and other industries to improve industry production and operational efficiency.

Digital people is a good business, but its large-scale implementation still faces difficulties in talent, cost, scenarios, technology, etc. Among them, the most critical one is the technical bottleneck. How to get digital people to get rid of the market assessment of "having a good skin for nothing" is a big problem.

However, with the emergence of large models, the development of digital humans seems to be ushering in new opportunities.

1. Large models empower digital people

For the industry, digital people themselves are not new. For a long time in the past, in order to strive for high-fidelity presentation in terms of digital human viewing appearance, the production cost was no less than that of making a professional-level movie.

A digital human research report released by UBS pointed out that the average initial investment cost of advanced virtual characters is 30 million yuan, and later a real team is required to complete shooting, dubbing, and editing. Taking Lehua Entertainment's virtual girl group A-SOUL as an example, the production cost of a single reached 2 million yuan, and the cost of an offline concert reached 20 million yuan.

However, this high cost problem does not solve the problem of digital human interaction effects. Due to its lack of intelligence, it is more like a soulless digital leather case than a digital human.

High costs, coupled with poor interaction effects, have limited the use of digital humans to experimental projects within manufacturers or large enterprise customers.

Therefore, as generative large models with learning capabilities take the lead in the content production paradigm, cheap digital humans targeting a wider range of small and medium-sized enterprise users and prioritizing large-scale implementation have become a feasible solution.

According to an artificial intelligence researcher, the reshaping and empowerment of digital humans by large models are mainly reflected in cost reduction and efficiency improvement.

From a technical perspective, building a digital human is mainly divided into modeling, driving, and rendering. Traditional digital humans mainly rely on computer graphics technology to capture real-person motion, which requires collecting a large amount of real-person data and in-depth modeling, which is time-consuming, low-efficiency, and high-cost.

Now, with the support of large models, through AI algorithms and based on deep learning models, action simulation, emotion simulation and other technologies, it only takes a few minutes of real-person video and several hours of training with large models to generate Realistic digital people, the production cost is greatly reduced.

Not only that, while the cost of digital human beings is reduced, the efficiency is also greatly improved.

Before the emergence of large models, digital humans had small differences in appearance, and could only answer questions "scripted" based on input unified scripts.

With the empowerment of large models, digital people have a "soul". Not only can their appearance and features be customized, but their intelligence and interactivity have also been greatly improved. For example, in some live broadcast delivery scenarios, digital people can already have basic interactions with the audience.

For example, the Xilin Digital Human released by Baidu Intelligent Cloud, with the support of large models, can quickly complete the construction of a live broadcast room in 15 minutes, automatically generate speeches that match product features, and start intelligent interactive Q&A.

In the live broadcast room of a certain catering brand, Xilin digital human anchors automatically generate live broadcast speech skills, including opening icebreakers, welfare broadcasts, warm-up speeches, order urging skills, etc. In this relay live broadcast between real people and Xi Ling digital people, the users did not notice at all.

What’s even more surprising is that in a real 6-hour live broadcast comparison, the digital anchor only needed 15% of the cost of the real anchor to obtain 85% of the GMV of the real anchor.

In addition to live streaming of goods, Baidu Intelligent Cloud Xiling, as the first digital human platform in China that fully reconstructs large models, can also provide enterprises with 2D real people, 3D realistic and 3D hyper-realistic portraits to realize video production , digital employees, digital human spokespersons and other applications.

For example, on the XiLing platform, it only takes 5 minutes of live video, half an hour to train a portrait, record 100 sentences, and generate an exclusive sound library 24 hours a day. Compared with the cost of live teaching, it only costs 30 in the past. %, recording efficiency increased by 20 times.

It is not difficult to find that digital people who have been reshaped and empowered by large models have to a certain extent got rid of problems such as high price and poor interactivity, and are increasingly appearing in short videos and live broadcast rooms.

At the same time, digital people are beginning to move towards more "identities" - bank financial planners, lawyers, teachers, deceased celebrities... Digital people are becoming anyone they want to be, and this is also the reason for this growing trend. A crowded track brings new opportunities.

2. Make digital people more like “human beings”

The emergence of large models has made digital people “reborn” and become one of the hottest entrepreneurial tracks today.

Currently there are two main categories of digital human manufacturers on the market: one is technology giants represented by Baidu, Tencent, Huawei, etc., which develop and launch digital human products based on their own advantages in large models; the other is technology giants represented by Baidu, Tencent, Huawei, etc. Small and medium-sized manufacturers represented by Silicon Intelligence, Mobvoi, and Xiangxin Technology.

Many investors in the field of artificial intelligence said that AIGC (generative artificial intelligence) is still in its early stages, and not many can be implemented to generate profits. Digital people are one of the few commercialization paths, and have already Profitable projects.

However, with the influx of a large number of entrepreneurs, industry competition has become fierce and homogeneous, and the track has gradually become crowded and involuted.

One manifestation of involution is that prices are falling lower and lower. At present, the price of most 2D digital people has dropped to the level of a thousand yuan, and some even only cost a few hundred yuan.

360 Group’s digital human marketing service also shows that it is promoted to customers through the supporting SaaS service of the intelligent marketing cloud platform. According to the monthly payment standards of member users, the price of a digital human is as low as tens of yuan. The maximum is around one or two hundred yuan.

With the influx of a large number of low-cost digital people into the market, digital avatars worth hundreds of yuan are quickly being hyped into a hugely profitable "wealth book."

Every late at night, when the mainstream platform is opened, a large number of digital people stick to the live broadcast room. Subsequently, doubts about digital people have gradually arisen, such as the effects are too fake, the market is chaotic, etc., resulting in poor user experience.

Some people in the industry believe that with the influx of more and more players, some entrepreneurs have limited technical capabilities and the quality of the digital human products they produce varies, which can easily lead to bad money driving out good money.

On the one hand, digital humans are mostly used in simpler environments to solve more basic problems. When switching scenes or facing multiple rounds of dialogue, they may answer questions incorrectly or fall into an endless loop, which restricts the user experience.

On the other hand, the interactive experience of digital people under large models is always significantly different from that of real people. For example, in the content generated by Sora, problems such as ignoring the principles of physics and the human fingers being more and sometimes less have been widely criticized. , which may further trigger the psychological "uncanny valley effect".

In this regard, some experts believe that it is necessary to further improve technological innovation capabilities and user experience, while ensuring the external anthropomorphic effect of digital humans, while optimizing the user interaction experience, focusing on real-time rendering, optical capture, three-dimensional reconstruction, Research and application of emerging technologies such as intelligent human-computer interaction and natural language processing, speech recognition, computer vision, generative AI and other technologies.

"The current technical goal that the industry is jointly breaking through is 'how to make digital people become more like a human being' and think like a human being," an industry insider said, accelerating technological collaboration between enterprises and solving The technical problems of interactive digital humans in aspects such as emotion perception and semantic understanding are the next focus.

This series of challenges all point to the technical side.

The 2024 "China Virtual Digital Human Influence Index Report" pointed out that as of the end of February 2024, the "Patent Search" data of the State Intellectual Property Office showed that the number of patent applications in the digital human field in 2023 was as high as 544 , reflecting the industry’s strong momentum and in-depth innovation in core technology research and development.

Judging from the institutions applying for digital human-related patents in 2023, the old Internet giants represented by Baidu and Tencent, leading communications and financial institutions represented by China Mobile and Industrial and Commercial Bank of China, Xiaobing Company, Leading digital companies represented by Shiyou Technology and Black Mirror Technology have formed a multi-power structure on the technology side.

Although the leading manufacturers in the industry have first-mover advantages in AI technology, in the process of rapid industry development, no company has absolute barriers to the technology itself. Essentially, people are behind the technology. For all digital human manufacturers, while AIGC brings opportunities, it also becomes the starting point for facing challenges.

3. Conclusion

Digital people are at the forefront, attracting a large influx of entrepreneurs who want a piece of the pie. But it is undeniable that as a young technology, digital humans are still in their early stages and the market is still being cultivated.

For digital human companies that are also in the red ocean, what they need to think about may not be how to "get low prices", but through the continuous advancement of technology, make digital people "more human-like" and improve the industry. The overall water level allows digital people to "survive" and create more value.

The above is the detailed content of After the blessing of large models, are digital people 'more human'?. For more information, please follow other related articles on the PHP Chinese website!