Source of this article: Times Finance Author: Xie Silin
Source: Screenshot of Tianmu Wensheng's sample video
Two months after Sora's explosion, A-share listed companies in China quickly followed up and launched a multimedia model with Wensheng video capabilities.
On April 28, the Tianmu model of Wondershare Technology (300624.SZ), a leading A-share digital creative software, opened public beta. According to Wondershare Technology, the model is the first audio and video multimedia model in China, and it has iterated nearly 100 audio and video atomic capabilities such as Wensheng music, multilingual dialogue, and one-click filming. Especially in terms of Wensheng video capabilities, the coherence of different styles, rich scenes and themes has been realized, and the one-click generation time supports about 60 seconds.
The generation time of this video can even be compared to OpenAI, the world's top large model company, released the Wensheng video model "Sora" in February this year. The product can generate videos of up to 60 seconds, astounding the world with its consistency, clarity, and ability to simulate the real physical world.
At the same time as the public beta, Wondershare Technology also announced its 2023 results, with significant growth in revenue and profit. According to the financial report data, the company will achieve operating income of 1.481 billion yuan in 2023, a year-on-year increase of 25.49%, and net profit attributable to the parent company of 86.2057 million yuan, a year-on-year increase of 113.20%.
Since the announcement on April 18 that it will be publicly tested the Tianmu model, it is not yet known how its video generation capabilities will be, but the stock price and market value of Wondershare Technology have been "moved". Times Finance statistics found that as of the close of trading on April 30, Wanxing Technology reported 92 yuan per share, with a total market value of 12.669 billion yuan. Compared with the closing price of 11.307 billion yuan on the 18th, its market value rose by more than 1.3 billion yuan, an increase of 12%.
Since its listing in 2018, the stock price of Wondershare Technology has gone through several ups and downs, as a "demon stock" that has been labeled as a number of concept stocks such as "unicorn" and "metaverse", is the Tianmu model of Wondershare Technology really as powerful as advertised?
After the public beta of the large model of the canopy, the outside world found that in addition to the indicator of generation time, the generation effect of the canopy is far from Sora. The performance in many indicators such as smoothness, coherence, and realism was not as good as expected.
Zhu Wei, vice president of Wondershare Technology, has publicly stated that there is indeed a certain gap between Tianmu's model capabilities and Sora. Especially in terms of Wensheng video effects, he once said in an interview with the media, "Sora, as an industry benchmark, is the goal that the company is trying to keep up." ”
On issues such as R&D investment and commercialization prospects, Times Finance sent an interview outline to Wondershare Technology.
Zhu Wei pointed out in an interview with Times Finance that in order to build a large model of Tianmu, the company is equipped with a R&D team of hundreds of people at the R&D level, which needs to invest tens of millions of yuan every year, in addition to the investment in data procurement and computing power, with a total investment of hundreds of millions of yuan.
In addition, Zhu Wei said that Tianmu has its own differentiated characteristics, compared with basic models such as Sora, Tianmu is more inclined to subdivide the vertical industry, hoping to focus on each specific application scenario, solve practical problems, and bring value to users. For example, AI singers, digital human broadcasts and other capabilities are embodied in specific subdivision scenarios, which are pure applications. "We hope that each product of Wondershare Technology is a combination of technology and application, and can solve specific problems in a certain segment. ”
According to incomplete statistics from Times Finance, none of the domestic Internet head companies that have entered the AI model have yet been able to generate more than 10 seconds of Wensheng videos, and more are picture videos. For example, Tencent Hybrid has cooperated with Tsinghua University and Hong Kong University of Science and Technology to launch the Tusheng video model "Follow Your Click", and the Alibaba Institute of Intelligent Computing has also launched EMO, which only supports the generation of singing videos with photos and audio.
So, as one of the first companies in China to claim to have 60-second AI video generation, what is the quality of Wondershare Technology?
The video generation is not as good as expected?
After the opening of the public beta, Times Finance made a trial application as soon as possible, ranking 1458th in the waiting sequence, and the application has not passed the review as of press time.
According to the information on the official website, Tianmu mainly has 7 capabilities, including text generation video, stylization of existing videos, text generation background music, video soundtrack, text generation sound effects, text generation pictures, and picture generation pictures. For example, Tianmu supports descriptors up to 2,000 words and offers two styles to choose from: realistic and anime.
Source: Screenshot of the official website of Wondershare Technology
However, the two Wensheng video examples provided by Wondershare Technology are only 2 seconds long, and the video images are rigid and single, with only simple camera movements.
On April 18, the 59-second sample video released by the official official account of Wondershare Technology was also stiff and slow. Compared to the long-form video that Sora released with one click, this video is very rudimentary in terms of consistency and coherence.
Huang Song, an AI algorithm expert and senior artificial intelligence practitioner, pointed out to Times Finance and Economics after watching the sample video that the video frame generated by the model is obvious, and there is almost no light and shadow effect, interaction between objects, and the cartoon effect is far from the real world.
A veteran who works for a well-known Tsinghua AI large-scale model startup commented to Times Finance that Wondershare Technology's canopy is "more like a stitching monster", that is, multiple short videos are stitched and edited into long videos of about 60 seconds.
In fact, the duration is the key link in the application of the Wensheng video model. Ai Changqing, an AI researcher at a major Internet company, pointed out that even if the video generation effect is good, if the scalability of the video duration cannot be supported from the underlying algorithm structure, it is difficult for the model to find a suitable landing scene, and it can only stay in the laboratory.
"The core concerns of the industry's video generation models include whether the video generation time is scalable and whether the generation quality is up to standard. The former is considered to be a fundamental and directional problem, while the latter is a problem of data quality polishing and training methods. Ai Changqing said.
On April 27, the day before the Tianmu public beta, Zhu Jun, chief scientist of Shengshu Technology and professor of Tsinghua University, released the video model Vidu at the Future Artificial Intelligence Pioneer Forum of Zhongguancun Forum. The model adopts the team's original Diffusion and Transformer fusion architecture, U-ViT, and supports one-click generation of high-definition video content up to 16 seconds and a resolution of up to 1080P. Vidu is not only able to simulate the real physical world, but also has rich imagination, with multi-shot generation and high spatiotemporal consistency.
Qiming Venture Capital, an investor of Biodigital Technology, commented: "Vidu is the world's first video model to make a major breakthrough since the release of Sora, and its performance is fully benchmarked against the world's top level, and it is accelerating iterative improvement. ”
Debang Securities Research Report also pointed out that in terms of the dynamics represented by the lens language, as well as the understanding of the laws of the physical world and the ability to simulate the laws of the physical world, it has achieved a similar level to Sora, and Vidu may play a role in the "catfish effect" of the domestic video model to encourage the breakthrough and innovation of the domestic multi-modal model.
It can be seen that there is still a certain gap between the ability performance of the Wondershare Technology Tianmu model and it.
However, there are also voices that believe that the domestic video generation model is still in its infancy, and it is understandable that the early results are not good. Ai Changqing pointed out to Times Finance that video is a space of more than three dimensions, and compared with text models, the training difficulty is increased by orders of magnitude, and it must rely on a large amount of computing power as support. At present, there is not enough computing power in China, and it is difficult to carry out research on video generation models.
"There is still a long way to go before the domestic video generation model is applied, and it will take at least two years or even more before there will be a model with better general effect. Therefore, Ai Changqing believes that it is actually a good thing that video generation models have been launched in China, and only by taking the first step can we communicate with each other and continuously iterate and improve our capabilities.
Times Finance also found that Wondershare Technology has recently attracted high-end talents in the field of AI with extremely high salaries. According to the 2024 spring recruitment information previously released by Wondershare Technology, the basic starting salary of AIGC-related product managers and R&D positions is more than 400,000 yuan, of which the monthly salary of key positions such as graphics and image algorithm engineers can reach up to 100,000 yuan, with an annual salary of more than one million yuan, and the monthly salary of multimodal audio and video algorithm engineers and basic platform architects is also more than 50,000 yuan.
Source: Wondershare Technology official website
Guo Tao, an angel investor and senior artificial intelligence expert, pointed out to Times Finance that for the emerging field of Wensheng video model, different companies' products may be at different stages of development, "Even if the Tianmu model of Wondershare Technology is not as good as Sora in all aspects, this does not mean that their technology research and development is worthless, and ultimately it depends on the combination with actual business and market application." ”
AIGC indirectly drove performance growth
For the vast majority of enterprises, the more important thing is how to invest their existing capabilities into business scenarios that can be quickly realized—how much business increment can the Tianmu model bring to Wondershare Technology?
As a listed company that sells creative digital software to C-end customers, Wondershare Technology has tried to apply AIGC capabilities to its products since 2023, not only its Wondershare Filmora Filmora Filmora PDF, Edraw Icon, Edraw Brain Map, Wondershare Easy Repair and other products have launched adapted AIGC functions, such as AI cutout, AI noise reduction, AI copywriting, AI image generation, etc., but also released the digital human short video marketing tool Wondershare Broadcast, AI virtual companion application TruMate, Innovative AIGC products such as Kwicut, an AI text and video editing tool.
But these explorations don't translate directly into revenue. Zhu Wei revealed to Times Finance that most of the company's original products have not been forced to implement commercialization strategies such as AI fees, so they have not brought direct revenue growth.
The impact is more indirect. Zhu Wei took Filmora as an example, saying that since the product launched functions such as text film and AI music generation at the end of October last year, it has driven the improvement of content service demand and user stickiness to a certain extent, and achieved a 100% increase in the subscription amount of asset resources during the year, further optimizing the revenue quality and efficiency of the product.
In its 2023 annual report, Wondershare Technology said that the introduction of AIGC capabilities, coupled with the impact of product iteration and upgrading, helped the company's document creative product line increase the number of paying users by 20%, with revenue of 130 million yuan, a year-on-year increase of 30.92%, the monthly active users of drawing creative products increased by 30% year-on-year, and the revenue reached 129 million yuan, a year-on-year increase of 30.52%, and the subscription renewal rate of video creative products increased by 5%, with revenue of 960 million yuan, a year-on-year increase of 28.91%.
However, with the continuous deepening of the application of AIGC technologies such as Wensheng Wen and Wensheng Tu, the product innovation space brought by it has gradually been exhausted, and the growth of Wondershare Technology has begun to slow down. In the first quarter of 2024, Wondershare Technology recorded an operating income of 358 million yuan, a year-on-year increase of 0.9%, and a net profit attributable to the parent company of 25.6498 million yuan, a year-on-year increase of 22.99%.
At present, the ability of video generation models that has not yet been applied on a large scale has become the key to the next growth of Wondershare Technology. However, considering the immature video generation effect at this stage and the still high training and inference costs, the path to commercialization is still unclear.
"In the field of large text models, free may be the consensus of everyone, because the cost is relatively controllable, and the more users, the more conducive to the rapid upgrade of large models. However, more must be considered for large video models with huge training and inference costs. Zhu Wei pointed out that in principle, Wondershare Technology will be more cautious in providing video AIGC capabilities to free users. At the same time, we will focus on pushing this capability to B-end users who are willing to pay and have relatively stronger capabilities, and truly push the audio and video multimedia model to commercial use as soon as possible.
According to Wondershare Technology's annual report, its top five customers accounted for 66.96% of its total annual sales.
Last year, I was asked twice if I wanted to hype up AI
Although the ability of large models is unsatisfactory, it does not prevent listed companies from "profiting" from market capitalization. Taking Wondershare Technology as an example, since last year, there have been voices questioning its hot speculation on the stock price of the AI concept.
Last year, when the explosion of large models drove the stock price of AI concept stocks to take off, Wondershare Technology has repeatedly released ChatGPT-related benefits on the interactive platform, such as saying that the company has a commercial account to open the API interface of GPT-4, has signed a cloud service framework agreement with Microsoft, and its main overseas products have been connected to Azure OpenAI services.
After some operation, Wondershare Technology also became one of the bull stocks that rose the most that year. Xueqiu data shows that in 2023, the share price of Wondershare Technology will rise from 30.65 yuan per share at the beginning of the year to the peak on June 20 of the same year, once touching the historical record of 172.2 yuan per share, an increase of more than 4 times. Then the stock price began to fall, and as of the end of 2023, it was still at 94.6 yuan per share, an increase of more than 208%.
Source: Pixabay
On March 27 and May 8, 2023, Wondershare Technology received two letters of concern from the Shenzhen Stock Exchange, asking about the specific situation of the company's related products using ChatGPT, AIGC and other technologies, whether it has brought direct income, and whether there are hot spots, speculation concepts and cooperation with shareholders to reduce their shareholdings.
Wondershare Technology replied that the company adopts AIGC-related technologies to achieve functional upgrades and intelligent efficiency improvements, and enhance the attractiveness of products. However, there is no separate charge for new AI technologies and functions, so it does not bring direct income, and it has not received the shareholding reduction plan of the actual controller, directors, supervisors and senior executives, so there is no hot spot, speculation and cooperation with shareholders to reduce their shareholdings.
However, less than half a month after replying to the Shenzhen Stock Exchange in May 2023, Wondershare Technology issued an announcement stating that Wu Taibing, the actual controller of the company, plans to reduce his holdings by no more than 2.06 million shares in the next 6 months, that is, no more than 1.5% of the company's total share capital. On July 26 of the same year, this shareholding reduction plan was announced, and Wu Taibing cashed out more than 200 million yuan based on the average price of the reduction.
It is undoubtedly a huge negative for the actual controller to reduce his holdings and cash out at the high point of the stock price, which undoubtedly casts a layer of fog on the AIGC of Wondershare Technology.
Before and after Wu Taibing's reduction in Wanxing Technology last year, there were also a large number of AI bull stocks that staged tricks to reduce their holdings, such as Kunlun Wanwei (300418. SZ), as well as 360 (601360.SH), which was questioned about divorce-style reduction, etc.
According to Hong Yong, an expert from the think tank of the 50-person forum on the integration of data and reality in China, judging whether a company's behavior is suspected of concept hype requires comprehensive consideration from multiple perspectives such as actual R&D investment, product functions, and market applications.
"If Wondershare Technology is only exaggerating in propaganda without substantive technological innovation or obvious product advantages, there may indeed be a suspicion of conceptual hype. However, if it continues to invest in technology research and development, and its products can be gradually improved and meet specific market needs, even if the current performance is not as good as the international advanced level, it should be regarded as a normal process of industry development. Hong Yong said to Times Finance.