Source: Number One AI Player

Author: Ah Hu

"Wait, I seem to see what the future child looks like!"

Recently, TikTok netizen @amanda_britany posted a picture scrolling video "I think they are really too similar".

In this video, she used Remini's AI function to predict the appearance of her child, and compared it with the photo of her child in reality. The similarity between the two shocked many netizens.

So far, this video has been played 4.4 million times, liked 295,700 times, and has over 100,000 favorites.

Relying on predicting the baby's appearance in the future, this application software named Remini topped the US application software list several times in July this year. According to data from AppFigures, an application analysis platform, Remini’s average daily active users exceeded 20 million in July, nearly doubling from the same period last year.

On TikTok, the topic #Remini# has reached 1.5 billion views, and more than 23,000 posts have participated in interactive discussions under this topic. Correspondingly, the topic #AIbabyfilter# (AI baby filter) also has more than 11.7 million views.

According to Remini's development team, Bending Spoons, in an interview with "INSIDER", Remini's AI function was updated in April this year, and within two weeks from July 2 to July 15, the app's global downloads were about 40 million times , the net revenue from in-app purchases in July was about 11.43 million U.S. dollars, and the estimated revenue was 7 million U.S. dollars (note: about 30% of Apple’s app market share was deducted).

Why does a software launched in 2019 have such a strong "long tail" effect? Is it easier to make a hit when developing mapping software on the AI application side? Let's take Remini as an example to see how the AI mapping application becomes a phenomenal product.

tens of millions of users, relying on Remini to predict the baby's appearance

For producers of indie games, the role of AI is greater than imagined. Remini was originally an image restoration editor developed by Beijing Dagong Technology Co., Ltd., a domestic enterprise. It uses image super-resolution technology to restore the details of old photos.

Out of consideration for data security, Remini has been managed and operated by Bending Spoons since Dagong Technology sold it to Italian mobile app developer Bending Spoons in 2021.

Compared with the image enhancement function of Remini before, Bending Spoons puts the focus of operation on the field of artificial intelligence synthesis. After the updated version, AI photo editing has become the focus of publicity on the Remini application page.

Remini adds more AI-related features

Open Remini, users can upload 8-12 photos of themselves or their partners, and then use AI to create personalized portrait photos.

The system will prompt users to use featured selfies, photos with the same theme, various expressions, and angles, so that it is easier to generate realistic and highly restored pictures.

After uploading the photo, the user can select the corresponding gender and template to generate the photo. Remini has more than 100 templates to choose from, including fashion photos, pregnancy photos, travel blockbusters, casual looks, and more.

Although it takes only 6 minutes to generate a picture for the first time, according to the actual measurement, the time to generate a photo after uploading 8 pictures is about 20 minutes, far exceeding the estimated time of the application.

Remini takes longer to generate images than actually predicted

In order to test the effect, I uploaded the photos of my parents, checked the child template, and wanted to compare whether the final generated effect was similar to when I was a child.

Remini will generate 6 pictures at a time. Overall, the generated photos can still extract facial features more accurately, but the details of teeth, hands and other details are not in place.

If you are not satisfied with the finished film, you can repeatedly submit selfies to generate photos, and you don’t need to pay for additional generation times. This process is a bit like opening a blind picture box.

Regarding the similarity of photos, friends around me have different opinions

Compared with artificial intelligence photo generation, this "predicting the baby's appearance" function is more like a face-changing synthesis in the later stage, capturing the essence of the face of the portrait, and then beautifying it into a childlike face.

However, Remini's model training samples may be mostly European and American homes, and the templates given are mostly based on children in Western countries. The generated pictures generally have the feeling of mixed-race babies, and to a certain extent lack Asian characteristics.

On social platforms, many netizens discussed their generated photos. It seems that Remini’s function of predicting the baby’s appearance is highly praised. Netizens generally believe that the effect of this function is better than other workplace photo functions. “The generated photos The effect surprised me, the similarity was more than 90%."

AI technology superimposed TikTok spread, bringing explosive growth

In the Remini application introduction, the official did not use the function of "AI Baby Generator" (AI Baby Generator) as a publicity point. The biggest "popularity" pusher is based on TikTok's fission communication, which brings a short period of time. The heat grows.

In July this year, a number of amateur bloggers on TikTok released experience videos of testing Remini’s functions. Most of the shared content revolved around pregnancy photos and children’s templates, which brought unexpected traffic to Remini.

Data.ai, an application product data statistics platform, shows that most of Remini’s open rate traffic comes from users’ natural search and direct access, accounting for 97.64%, while advertising marketing only brings 0.6% of Remini’s traffic.

Image source: application product data statistics platform data.ai

The ranking growth curve of Remini's downloads in various markets almost coincides with the search volume of the keyword #Remini# on Google search engines and TikTok platforms, which can also explain why Remini did not have explosive growth in the early stage, but downloads appeared in July A case of soaring volume.

The majority of netizens have always had a strong curiosity about predicting the baby's appearance and spying on the future aging appearance. According to a previous user survey of the face editing software FaceApp, more than 80% of users have used the "AI aging filter" in the application. This function satisfies their curiosity and desire to explore the future.

Probably for similar reasons, some US technology innovation companies have successively launched corresponding "prediction" applications since 2018, such as "BabyMaker" and "Future Baby Generator".

"AI baby appearance prediction" applications emerge in endlessly

In China, the need to "predict the baby's appearance" is not uncommon. Previously, small and medium merchants on the Taobao platform used AI drawing tools Midjourney and Stable Diffusion’s matting function to post-modify through 4D color Doppler ultrasound photos to generate baby looks. Although this business looks like an "IQ tax" to many people, there are many people who are willing to pay for it.

Many users who share "baby looks" on social platforms are novice couples who have just become parents. I want to use Remini to prove "who is the newborn child more like?".

TikTok user @kjmnailz wrote in the video copy that "father's genes won" and attached pictures generated by Remini software and photos of her daughter. She said that she initially just wanted to see how accurate the photos generated by AI were.

In addition to effectively grasping the topic of predicting the baby's appearance, the open source update of the underlying model (such as LoRA, StyleGAN) also contributed to the outbreak of Remini.

Since Remini did not disclose the basic model behind it in public information, the current technical analysis is based on the industry's empirical judgment on existing AI mapping products. According to the analysis of the creator of the Stable Diffusion model, Remini may have specially trained a deep learning model. After specific parameter training, the AI model can learn the correlation between adult facial features and existing data sources, and then perform alignment and matching. That is to beautify and process the photos input by users, so as to achieve the effect of "predicting" the appearance of future children.

In addition, a professional on the product side of an AI technology innovation company speculated that Remini used the AI technology framework of StyleGAN, which can create realistic and high-resolution photos by combining the features of two images, or used a lightweight The LoRA plug-in assists in the layout of mobile products.

"George Zhou Shicheng", a practitioner of film production and game engines, and the technical up master of station B, told "Top AI Player" that the underlying technical logic of such AI mapping tools is very similar, mainly to achieve the effect of "changing faces". Players can also use AI plug-ins to perform basic synthetic face changes, but the corresponding details need to be trained by themselves. If it is a video face change, you need to choose a different AI tool.

It can be seen that the development of these AI mapping applications is not difficult, at least there is no major bottleneck on the technical level. The difficulty lies in how the development team can accurately grasp the user's psychology and combine product functions with actual needs. Stand out from the crowd of apps.

**Can Remini's success be replicated? **

In addition to Remini, the previously popular Miaoya camera is also a phenomenon-level product in the field of AI mapping. Is it easier for the AI mapping application to release explosive models?

With its strong photo processing functionality and fast processing speed, Remini successfully attracted a group of users at the beginning of its launch, which also helped Remini enter the Middle East and Southeast Asian markets such as Saudi Arabia and Indonesia, and accumulated users from all over the world.

The reason why it can succeed is that it already has a huge user base and has core functions related to mapping. It can quickly use social hotspots to spread, so that retained users can generate re-consumption needs.

In addition, social platforms such as TikTok have also become one of the key channels for application software to acquire users. After viewing the short video, users will ask how to achieve such an effect in the comment area. They have a certain curiosity and freshness about the novelty material, and are willing to download and experience it. The conversion rate is also relatively good. Therefore, more and more AI mapping applications have become popular with the help of TikTok.

However, at present, many AI application start-up companies rely too much on a single AI function and social communication to achieve commercialization. This model of relying on short-term explosions to acquire users may make it difficult for products to obtain sustainable value. For example, the previous FaceApp, whose main function is the AI aging filter, also relied on social communication to gain a lot of attention, but most of it was a short-term growth effect. In the long run, the commercialization effect is limited.

In contrast, relying on the core mapping function to obtain users, coupled with AI as an auxiliary embellishment, this kind of commercialization logic is more stable. Even if the popularity of a certain AI function declines, such companies can continue to introduce new AI special effects or filters, and will not be as prone to falling into growth difficulties as a single AI application.

"In the field of AI, freshness is extremely important. Especially for young users, if the application is not updated in time, the possibility of user loss will be greatly increased." Dagong Technology CEO Huang Shuo said frankly in an interview, "Back to the essence of technology , focusing on product innovation is equally important.”

The development team combines the new AI technology to realize the application in a short period of time around the general needs of users. In this case, it can usually obtain better technology and traffic dividends.

"Actually, in the field of AI mapping, I personally think that it may be easier for laymen to make money." "George Zhou Shicheng" believes that "Practitioners who are really engaged in AI mapping look down on AI. The current AI image generation technology still has certain limitations. It can only be used as an auxiliary tool. But this "limitation" can already meet the needs of most players, and many experts are rejecting AI. In fact, laymen can find some new business opportunities from it."

In his view, the next opportunity for AI mapping should be in the field of 3D rendering, such as using AI renderers in 3D design. It will also slowly "roll" into professional fields, including architectural interior design, scene concept design, game or film and television production, application development, etc.

For ordinary users, it is difficult for most people to make an AI plug-in at low cost, copy and apply similar functional effects. Professional technical knowledge such as code compilers and function deployment has blocked many consumer groups.

The main opportunity for more ordinary players at this stage may lie in how to play some new tricks on AI mapping, such as creating novel games such as "predicting the baby's appearance" and "embedding text in pictures", using social media to form hot topics, and then evolve For a new trend.

The gameplay of AI is changing with each passing day. If there is no way to grasp the technical dividend, then firmly grasp the traffic outlet.

"Some players used AI drawing to make avatars and publish tutorials, which is actually very easy to realize. If AI drawing is gradually 'rolled' into the professional field, it will be difficult for ordinary people to make money from it," added "George Zhou Shicheng".