Meitu’s self-developed large model 3.0 has been officially released!

And it is fully used in Meitu’s imaging and design products.

Picture

This is the latest iteration of the Meitu model 100 days after its release.

Compared with the original version, version 3.0 can generate more realistic and delicate picture details.

Picture

These generation abilities can be directly experienced on Meitu XiuXiu.

Picture

The AIGC gameplay that is popular recently can be found in it.

Picture

Meitu’s founder, chairman and CEO Wu Xinhong revealed that most of Meitu’s products have been integrated with its own research and development Large models

In addition to the imaging and design fields, Meitu's self-developed large models will also develop in the five major industries of e-commerce, advertising, games, animation, and film and television.

Meitu Xiuxiu can experience it directly

Meitu’s self-developed large model is called MiracleVision (Qixiang Intelligence).

The most notable feature is that it understands aesthetics better.

Meitu integrates its long-term accumulated aesthetic cognition into the MiracleVision large visual model, and establishes an aesthetic evaluation system based on machine learning to assign "aesthetic scores" to the results generated by the model to continuously improve the model. Understanding of Aesthetics

So Meitu also calls its large models “a combination of art students and science students.”

Currently, the latest capabilities of MiracleVision3.0 can be experienced on Meitu XiuXiu.

Picture

The operation is very simple. For example, using graffiti to generate pictures, you can easily create them with just a few simple sketches and prompt words. Complex models

Pictures

On the AI ??visual creation tool "WHEE", you can get a more professional experience and be able to control more functions

For example, in the "Picture to Generate Picture" function, we can use intelligent association to generate more complete prompt words

You can control the content you do not want to present through text

Picture

You can further control the generated image style.

Pictures

And through parameter adjustment and screen control, the generated pictures can be more in line with your expectations.

MiracleVision, a large model of Meitu, has now set foot in five major fields: e-commerce, games, animation, advertising and film and television, aiming to improve the efficiency of workflow

Picture

MiracleVision has undergone three updates since its first release in June

• The 1.0 version mainly builds the large model architecture and foundation;
• The 2.0 version unites designers, Teachers and students from art schools jointly constructed high-quality data sets, allowing the large model to reach a better state in terms of aesthetics; version 3.0 of
• emphasizes the controllability of the model, improving details and local control. This enables the model to achieve better results at the technical level.

Pictures

Currently, the number of AI effects generated by the Meitu AI visual creation tool "WHEE" has exceeded 5.5 million and has been initially created. Designer ecology; the number of monthly active users of the AI ??spoken video tool "KaiShoo" has exceeded 350,000; the desktop AI video editing tool "WinkStudio" has cumulatively served nearly 100,000 video creators.

Meitu Wu Xinhong also said that currently all Meitu employees are actively embracing AI. With the integration trend of artificial intelligence and image computing, Meitu has formulated a clearer development strategy

, such as quickly launching self-developed large models and deploying five major industries. These actions are based on Meitu’s understanding of AIGC trends and Insight.

Invest 300 million in R&D in the first half of the year

The most obvious aspect is Meitu’s understanding of the current AIGC trend of active users

Wu Xinhong said that they have observed that nowadays The people closest to AIGC trends are not designers, but “communicators”.

Mainly focused on new media operations, e-commerce operations and KOL groups

do not have the "baggage" and high requirements of professional practitioners, and at the same time can downward Connecting this group of more ordinary people is one of Meitu’s current strengths. They are gradually developing more complete AI tools to build workflow

Picture

Meitu also believes that under the trend of large models, it has a The natural advantage is that it can quickly obtain feedback from users

As a product company that has been established for 15 years, Meitu has built a large enough user group to be able to test new features on its own platform. , quickly get market feedback and quickly adjust the effects, so that large models can be more efficiently combined with actual scenarios.

As a provider of AI capabilities, Meitu Wu Xinhong believes that the current large AI visual models face three problems to be solved on the production side: ultimate effects in vertical fields, workflow integration and monetization capabilities

Looking at the one-to-one correspondence, Meitu has begun to make efforts in these three aspects.

First of all, Meitu improved the model generation effect within 100 days through rapid iteration and self-developed large models, making it richer and more controllable. On the basis of satisfying users' fresh experience, Meitu can provide image design works with more practical value

Secondly, Meitu has chosen five major directions to provide AI design tools for producers. According to reports, Meitu chose the e-commerce, games, animation, advertising and film and television industries mainly because they have a more native relationship with the Internet industry. This allows AI tools to be better integrated into their work links and integrate workflow faster.

Third, liquidity. Meitu has currently built a creator ecosystem. These creators who are interested in AI can make their own original AI visual effects into formulas and sell them on the platform built by Meitu. In this way, creators can earn more income, and they can also play a role in promotion and attract more people to use AI visual creation tools.

Based on these understandings, this year Meitu officially proposed the form of AI products, including the bottom layer, the middle layer and the application layer.

The bottom layer is a large self-developed model by MiracleVision.

The middle layer is composed of Meitu AI open platform, Meitu creative platform, and Meitu AI assistant RoboNeo, which are respectively responsible for opening and sharing AI technology, building a creator ecosystem, and aggregating Meitu image services

The application layer consists of various types of application apps that are more familiar to ordinary users.

The purpose of building a three-tier architecture is basically from the perspective of product

Take the underlying research and development as an example. Meitu Imaging Research Institute (MT Lab) under Meitu launched the Established in year. At present, the company has about 600 people related to large model R&D. In the first half of this year, the company invested 294 million yuan in R&D, a year-on-year increase of 10.8%.

Meitu stated that they follow a more pragmatic logic when conducting research and development, that is, taking into account whether they can actually grow users and revenue. At the same time, they believe that all this can be proven through actual performance

In the first half of this year, Meitu achieved total revenue of 1.261 billion yuan, a year-on-year increase of 29.8%. The number of monthly active users reached 247 million, a year-on-year increase of 2.5%. The number of VIP members exceeded 7.2 million, reaching a record high. Revenue from the imaging and design products business, mainly VIP subscriptions, was 602 million yuan, a year-on-year increase of 62.2%; revenue from the beauty industry solutions business was 286 million yuan, a year-on-year increase of 31.1%; advertising business revenue was 349 million yuan, a year-on-year increase of 28.2%; other businesses Revenue was 23.09 million yuan.

But back to the present, Meitu also admitted that the application of visual large models is still in the exploratory period. At present, many application developments are going hand in hand, and each company is also exploring its path.

Wu Xinhong believes that the popularization of the application of large visual models will go through three stages: the exploration period will be before 2024, the rapid development period will be from 2024-2025, and the mature period will be from 2026-2030

PictureTo successfully cross the cycle, you need to have the ability to examine trends, and you need to have real financial support

Let’s use the current Manufacturers are worried about the computing power issue. Meitu revealed that it is currently cooperating with a number of cloud vendors and has sufficient computing power to use.

Wu Xinhong further added that Meitu is lucky to be able to earn rewards through subscriptions and individual purchases

The above is the detailed content of Meitu Xiuxiu has launched the latest self-developed large model, and you can directly experience various AIGC gameplays.. For more information, please follow other related articles on the PHP Chinese website!