Authors | Wan Chen, Zheng Yue
Editor | Zheng Xuan
No one expected that the first truly groundbreaking product in the domestic AIGC would emerge in the field of photo beautification (realistic portraits), which is already quite mature.
Quietly launched in mid-July, with the gimmick of "only 9.9 yuan and 20 photos can generate artistic portrait-level personal beauty photos," Miaoya attracted countless beauty-loving girls to download and try it out in just a few weeks, and the generated "Miaoya beauty photos" swept through social media such as WeChat Moments.
Miaoya's breakthrough speed is eye-catching, but the development team behind it is very mysterious. Before going online and quickly breaking through, there were no spoilers at dozens of large model release events in the past few months, nor did it start with tens of millions or even hundreds of millions of angel financing to "start a business after leaving a big company." Even the vast majority of insiders are not clear whether Miaoya is a startup team or an internal project of a big company.
Recently, Miaoya's product leader Zhang Yueguang accepted a group interview with many technology and business media, including Geek Park.
Miaoya is an internal project incubated by Alibaba's Big Entertainment. After ChatGPT set off a new wave of AI, Zhang Yueguang and a few colleagues formed an internal group called "AIGC Breaking Group" in February this year, and in one month, they discussed the direction of the "realistic portrait" project, and then spent three months polishing the specific product "Miaoya."
This is reminiscent of Google's famous "20% free time system" in the early years, and the amazing innovations such as Google Glass, Cardboard, and self-driving cars that were born from it. This is more like the romantic opening of Silicon Valley-style innovation, and it has also aroused our greater curiosity about Miaoya.
In a one-and-a-half-hour interview, Zhang Yueguang answered a series of questions and controversies about Miaoya. Although they classified issues such as the technology behind the product, user scale, team composition, and future commercialization strategy as "inconvenient to disclose" due to commercial secrets, they provided detailed explanations for the reasons for choosing realistic portraits, the process of product birth, data privacy controversies, and the reasons for its popularity, revealing a glimpse of the mysterious veil of Miaoya.
The following is the conversation content organized by Geek Park.
A small interest group of several people, polished "Miaoya" in three months
Q: How did the team start from 0 to 1?
A: The project started in January and February, it was relatively simple, and there was no clear business plan like now. It was similar to a small interest group, where I and five or six classmates studied cutting-edge knowledge and products in AI and image technology. On February 9, I formed a group called AIGC Breaking Action Group. From then on, we gradually paid attention to some new technologies and tools in the image AI technology field. Around March, I thought that some key technical elements had become relatively mature and decided to develop in the direction of realistic portraits.
Miaoya team photo | Miaoya Camera Official Account
We spent more than three months polishing this direction, including templates for technical and aesthetic projects. By the end of June and early July, we finally felt that we could show its status to users, started a two-week internal test, invited some friends and colleagues, and received feedback from everyone. We found that users were quite satisfied with the product's effects. In mid-July, we officially launched the product to the public. After the product went online, it received a significant response and users indeed liked it, and its development speed exceeded the team's initial expectations.
In terms of resource support, the initial resource consumption was not large, and as it gradually developed, Miaoya received more support. Most importantly, the company's tolerance and support for innovation allowed us to make many attempts.
Q: Why choose the field of realistic portraits for AIGC landing?
A: I have some basic analysis of the product direction. First, I have done album-type products before, which provide users with algorithm-based automatic classification. Among all user albums, over 70% are photos of real people. If all photos and image information are put together without screening, realistic portraits will be the most valuable, with the highest proportion and the most attention.
Another consideration is that overseas products like Midjourney have been relatively successful, and there are also some similar products in China trying to replicate this model. However, these products themselves are like a painter who can draw anything, and they have relatively high requirements for technology and algorithms. More vertical directions are easier to achieve, which is why we chose the broad direction of realistic portraits.
Q: Many people who have experienced Miaoya say "only 9.9 yuan can beat Tianzhen Blue and Hippocampus," what do you think?
A: We did not intend to defeat Tianzhen Blue or Hippocampus, this is just a slogan from users. The initial form of our product may be similar to the current photography service form, but from our perspective, we hope to reach a consensus with the photography and videography industry and become an enabler for the industry. Hippocampus, as a physical photography service, can never be replaced. Because the process of shooting itself is very important for customers. For example, when taking a family portrait, the process of the family gathering together to complete it is a very meaningful thing. Even if AI technology is more advanced, it cannot replace such a process.
Miaoya Camera homepage
We hope to cover users in some relatively underdeveloped cities and regions, because they may have difficulty accessing such services. However, the demand for beauty and photography exists everywhere, because people always want to leave behind memorable photos for themselves.
In our initial thank-you note, we mentioned: we invited photographers and designers to participate in template creation, and their help with AI technology was also amazing. We believe that we are complementary to this ecosystem, rather than in opposition; we want to provide some tools for this industry, rather than eliminate competitors.
Q: Looking back, why did Miaoya become popular?
A: The most important factor is the determination and hard work of the team members. The team atmosphere is simple, direct, with clear goals, and pursues an ultimate style. We spent more than three months polishing an effect to achieve a better user experience and effect. The second is the company's encouragement, support, tolerance for innovation, and further support after seeing some phased results.
If we look for reasons in the product itself, I think there are several points.
First, I think we were lucky. Second, user experience is very important. If users cannot use the product smoothly, then even if the product itself is of high value, it will be abandoned by users. The commercial value of a product is not a linear relationship, but only after reaching a certain threshold can it truly be reflected. Therefore, when designing a product, we must pursue high-quality effects, at least reaching a score of 90 or above. When the effect reaches a point where everyone in the team feels that the effect is really good and there is a wow moment, then its probability of viral spread is very high.
In addition, the controllability of the product is also very important. When designing the product, we insist on the "real, like, beautiful" three things, "real" has no AI sense, "like" is like the user, "beautiful" is 3 points more beautiful than the user. These requirements all need to reach a quality of 90 or above, and only then will we push the product to the market.
Q: Why did you set the number of photos uploaded by users at 20?
A: To achieve the ideal effect, 20 photos are the best solution at this stage. It's not convenient to say more specifically.
Miaoya camera requires users to provide 20 or more photos | Miaoya screenshot
Q: Some analysis suggests that popular tool-type products like this will face the crisis of being copied. Some people speculate that our model behind the scenes is based on fine-tuning open-source models, and believe that the technical barriers of such products are not very obvious. From an internal perspective, where do our technical barriers or product barriers manifest?
A: It is indeed inconvenient to disclose the specific implementation of our technology, but we do have some barriers. The most important thing is the continuous iteration and expansion of the subsequent effects, as well as the continuous polishing of the effects by our technical team. We will continue to optimize the product's effects.
Q: Some people speculate that Miaoya used the Stable Diffusion model and then the Lora fine-tuning method, does this speculation reflect reality?
A: This may indeed be inconvenient to comment on. Any questions about the technical implementation may belong to our trade secrets and may not be easy to disclose.
Q: Was the base model for image generation behind Miaoya trained by yourselves? Will you use Alibaba Cloud's image generation model?
A: Miaoya is an internal project of the Alibaba Big Entertainment team and has not used some of the technologies related to Alibaba Cloud. As for how it was implemented, I can't really talk about it.
We have a model called "Tiziano," and you should have seen this name in the (official thank-you note released on July 17) — the name of the father of portrait painting. Through this name, you can also see that from the beginning, we wanted to provide users with realistic portrait services.
Q: How did the current style of Miaoya's templates come about?
A: There are many young students in the team, and everyone discusses efficiently. It's more about looking at users, such as in cases where they are publicly spreading on social media and WeChat Moments, the directions that users prefer.
We also collect a lot of user feedback, such as what kind of templates users like, and continue to launch templates based on user demand. The core is to meet user demands. The upcoming batch of templates is basically generated based on user voices.
Q: How does Miaoya's camera balance the realism and beauty of the photos?
A: Seven or eight parts like, two or three parts beautiful. When we set the business direction for realistic portraits, we mentioned the three directions of real, like, and beautiful. One is without an AI sense, then it should be as much like me as possible, but more beautiful than me. Whether it's a beauty camera or every generation of shooting products, this is common sense.
At the same time, we give this choice to the user. When the portrait first comes out, it is very beautiful, but it may not be very like the user, and then we have a special function that can make the photo more like me. Click it a few more times, and it will become more like me.
Q: Will Miaoya always target young female groups?
Answer: The main users of all selfie or image-related products are female users, which can be easily seen from the number of our templates.
Privacy and Data Security Controversy
Q: With the backing of Alibaba's many film and television resources, do you have plans to develop products for dynamic face swapping? For example, directly swapping a user's face with a character from a movie or TV show.
A: We do not aim to collaborate with Alibaba's Big Entertainment, but if there are suitable opportunities, we can consider collaborating with film, ancient-style, and palace-themed products to understand user interests.
Face swapping itself is something we are determined not to do, because from the beginning, my definition of the product is different from previous face-swapping products, and it's something I deliberately don't want to do.
Q: How do you view the development of face-swapping products on the market?
A: Face swapping is a fairly mature product, and I won't comment further on its product scenarios. On the other hand, it also has certain security issues.
The photos we generate using AI are actually less realistic than direct face swapping, at least we have to pass the face key point detection security level, it just looks like it. Our approach ensures the security of user data and the relatively safe generation of content, providing more long-term development potential.
Q: How do you view the controversy over privacy terms?
A: After Miaoya went online, there was some controversy due to the user agreement, and we made modifications immediately. We acknowledge that this was our own problem. From a certain perspective, the agreement was not fully explained, or it did not use the simplest way to inform users.
Even the old agreement, if carefully read, would reveal that the rights to all generated content belong to the user, and we cannot use it for secondary purposes. But to provide normal services to users, we need to display and manipulate the user's work in the product, provide sharing and downloading services, and definitely need to obtain user authorization. The original agreement also meant this, but the wording may have been problematic, so we made modifications.
Our intention is that everything we do belongs to the user, and we have no ownership. We don't keep any of your data, and we won't use it. What needs to be displayed, we'll show it to you, and that's our original intention.
Q: When using Miaoya's products, users need to upload 20 photos, but you mentioned that the platform will delete the user's digital avatar data after the photos are generated. Why did the previous version mention using the data for training?
A: First of all, we have never used this data for training, and we never will. The previous agreement scared everyone.
Once again, the previous agreement's previous clause states that all rights to the generated content belong to the user. Because we completely transfer all rights to the user, we need to obtain the user's consent and authorization to display this information.
In fact, if you pay attention to the user agreements of many products, they do not state that the ownership of the items created on the platform belongs to the user, it's vague, or even belongs to the platform. If it belongs to the platform, then there is no need to request user authorization, as the platform already has ownership.
This operation is definitely our problem, as it turned our original good intentions into a very scary form. Here, I must clarify once again the original intention of our team, we have never considered the misuse of user privacy data, so when we found that the wording was not user-friendly and not understandable to users, we immediately made modifications, and we felt no pressure at all, because we never did that in the first place.
Q: Will users' facial features be extracted and leaked?
A: We have not extracted the user's facial key points, it's not something we do. When users upload photos, there is a recognition process, which involves some security-related checks to avoid users uploading photos that do not comply with national legal requirements, but we do not obtain the user's facial data points.
Secondly, the process of creating a digital avatar does not involve extracting facial features. In fact, we cannot reverse-engineer the 20 photos you uploaded. Although it's not convenient to disclose the specific technical principles, it does not involve the extraction of user facial data information.
This is also why, as I mentioned earlier, if you post your own face on social media or even on your friends' circle, the risk of this, if you are concerned about security, may be higher than our generation-type products.
Q: If you are not extracting the user's facial key data, how do you achieve the current effects?
A: It uses some current AI technology, but it definitely does not involve extracting facial key points. I believe many people in the industry will have various guesses about the technical route, and I cannot directly confirm it, but it definitely does not involve security issues. Everything in AI is probabilistic, it cannot replicate your photo and reason it out, so it is very safe.
Q: In terms of protecting user privacy, what measures has Miaoya taken?
A: In the agreement, we promise not to retain user photo data. All photos uploaded by users, whether 20 or more, will be deleted after the digital avatar is created. So we have not retained user data, and there is no way to reverse-engineer this data from the user's digital avatar.
Secondly, we will not show your digital avatar or any other data to third parties. In our product form, each user can only see the images they have generated and can choose whether to share their creations.
Thirdly, we have also implemented a lot of network security protection. Because even if we protect user privacy and do not misuse it, we also need to prevent third-party hackers from obtaining it, but this is thanks to Alibaba's overall technical support in privacy protection and security.
I hope everyone does not have excessive fear. This tool is essentially a graphics tool, and the content created can only be seen by you, and it does not have the attribute of public dissemination, making it even safer. Some may ask, since AI can generate my face, can it be used for malicious purposes? But in reality, no matter how it is generated, it is just like you. But if, for example, someone directly takes your face from a public social media platform, then that's you, and they can do bad things. On the other hand, the images we create, which are similar to you, are a product with less risk.
Overall, this is a product with very low impact on users, almost none. Data is not retained, your data cannot be reverse-engineered, we have not obtained the original data and key points, everything can only be seen by you, and whether to disseminate your generated data is your personal choice. Even if the generated content is disseminated, it cannot be used for security checks, because it is just like you, but not exactly the same as you.
The First Day of AIGC Products
If you don't charge, you won't receive money
Q: How are the download numbers and user retention of the Miaoya app now?
A: Specific data metrics are confidential and cannot be shared. But we believe the data performance is good and meets expectations. In the two weeks of internal testing, there were approximately 10,000 users, mainly from colleagues, and most were from the internet industry.
Regarding the future user scale, we have our own plans, but cannot disclose them. We hope to provide this kind of product or service to a broader range of users to meet future development trends.
Q: How does Miaoya's camera maintain user engagement? Some voices in the market are saying that this type of app is inevitably short-lived, what is your view on this concern?
A: First of all, the phenomenon of effect-based products going viral is almost commonplace in this industry. Every few years or during the peak of the internet, similar products appear every year. Some products do attract public attention and then disappear, but some will settle down. For example, the product Lian Meng, everyone thought it disappeared after going viral. But in reality, the demand for Lian Meng was eventually replaced by the special effects and filters of Douyin. This is a very stable demand, and there are still many users using it every day.
For Miaoya's camera, we certainly hope it is not a "one-hit wonder," but a stable service demand. Of course, this requires us to continuously improve the product's effects and service quality, and to continuously introduce more services. The specific implementation method cannot be shared, but our goal is to provide users with low-cost, high-quality image services.
Q: Miaoya's B-side workstation will go online on August 4th (today), and has invited some designers who have performed well in AI and related technology fields for internal testing. Does this workstation have content community attributes? How will it collaborate with photographers? What kind of designers is it mainly attracting? What are the core pain points of the B-side? What can Miaoya solve?
A: The B-side workstation does not have content community attributes. Our current ToB business is relatively simple, more like a supplement to the ToC business. We mainly hope to collaborate with designers and photographers who have research and interest in AI technology, provide them with services, and work together to create templates to meet user needs. The specific form of collaboration cannot be disclosed, but in simple terms, it is to create templates together.
Q: How did you come up with the idea of selling it for 9.9 yuan? Why did you start with commercialization from the beginning?
A: There are two elements to the commercial potential. First, the cost of computing power is expensive, as everyone in the industry knows. Secondly, AIGC and the business logic of the internet era are fundamentally different. In the AIGC era, if you cannot start charging users on the first day, you will never receive money from users. AIGC products should start charging users from the first day of their birth.
If you abstract all internet businesses, they are essentially a form of information flow and channel business. The platform's role is to move and distribute information from point A to point B. Regardless of the information being distributed, abstractly speaking, it is a channel. Initially, this channel is free, but gradually increasing the commission or through other means, it achieves secondary monetization.
However, what the AI era changes is not the channel, but the factory. AI products themselves are the factory, not the distribution platform. If the products produced by the factory cannot be sold on the first day, they will not be sold in the future. This is why I firmly believe that good AI products have great potential for commercialization.
The reason Miaoya chose 9.9 yuan is after various calculations and considerations, hoping to provide a relatively simple and understandable price point. From the user-driven dissemination we have observed, the price point of 9.9 yuan is relatively easy for users to accept and understand.
Q: After the surge in popularity, the queue time was very long, currently it has been shortened but still at 3 to 4 hours, what optimizations have you made? What are the costs?
A: The current reduction in queue time is the result of a significant increase in computing power, but while we are increasing capacity, user usage is also increasing, so the queue problem will still exist. The essence of solving the problem lies in expanding computing power. The cost of expansion is controllable, and there are many factors to consider when pricing. I have communicated with other teams, and they have indicated that there are many hidden barriers in developing this product that cannot be overcome. On the surface, other teams could also develop similar products. However, they did not develop it because if there are too many users uploading, the cost of computing power would become particularly high, and economically it would not be cost-effective. The company has given Miaoya a lot of space and support in this regard. With the addition of payment, we are relatively healthy in this matter.
Q: After the ToB product goes online, will there be other pricing besides 9.9 yuan?
A: Please continue to pay attention to the product, it's not convenient to discuss specifics. It definitely needs to comply with basic business logic, ultimately needing to make money, considering revenue, costs, and profits. It's not convenient to calculate whether it has already achieved profitability, but I can only say that I believe it is a relatively healthy business model.
Q: To what extent do you think the realistic portrait track has been achieved, and when would you feel satisfied and ready to explore other businesses?
A: I think the road ahead is still very long, we are just getting started.
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
