Image source: Generated by Unbounded AI

In 2012, computer scientists Andrew Ng and Jeff Dean conducted an experiment.

They used 16,000 CPUs, 10 million cat pictures and the world's largest deep learning network at the time to train for 3 days and generated the world's first AI cat face image. This was the beginning of people automatically generating pictures based on deep learning models. Due to the technical capabilities at the time, the kitten in the picture was “unrecognizable”.

Twelve years later, when netizens saw the picture of a black cat with white wings for the first time, almost everyone thought it was a real photo.

After all, it is not unusual for pet bloggers to decorate their cats with wing accessories and then use a film camera to take photos and record them. It wasn't until they saw the "AIGC" tag at the bottom of the post that people were surprised to find: "This was actually generated by AI?"

AI kitten 12 years ago vs. AI kitten 12 years later

With great curiosity, more and more netizens are following the Internet cable to an application called "Recraft AI" to try to generate their ideal kitten.

The situation is getting out of hand.

In recent days, all kinds of "kitten film photos" have filled the information flow of social platforms such as Xiaohongshu and Douyin. Whether it’s a kitten holding a red wine glass, a kitten typing on the keyboard, or a kitten holding a magic wand casting a spell, they are all masterpieces of Recraft AI.

Of all the developments in the field of AI in recent years, the one closest to ordinary users is AI-generated pictures.

Since OpenAI released the DALL-E large model at the beginning of 2021, AI has been able to generate images through text. The AI ​​image generation tools that were born within three years have evolved and iterated in waves. Those specializing in breakthroughs in technical capabilities include Stable Diffusion, Midjourney, FLUX, etc., and those specializing in C-side applications include Miaoya Camera, Remini, etc.

From the initial pursuit of the ultimate "image", to now gradually getting tired of the taste of AI and starting to pursue style and aesthetics, this generation of netizens are putting Recraft AI on a new "altar".

"Dream AI" without AI flavor

This time What became popular was actually “Hard Flash”, a preset style built into Recraft AI.

This mode can simulate the shooting effect after turning on the flash when taking film photography. The resulting picture will have a prominent subject, high contrast, and full and rich colors. Currently, the AI ​​website allows free users to earn 50 points and generate 50 pictures every day. Although the generation effect of Chinese prompt words is still not as good as English, Recraft supports direct input of Chinese passwords.

If we say,The emergence of ChatGPT makes the author anxious, and the emergence of Midjourney makes the painter sad. This time, it is the photographer who is panicking.

Xiaohongshu is where Recraft first attracted attention from domestic users. Currently, there are more than 10,000 notes on related topics.

There are many photography enthusiasts lamenting, "I declare that photography no longer exists"; there are also professional film photography bloggers who, after viewing the pictures generated by Recraft, think that the composition, color and aesthetics of the AI ​​are extremely Good, and began to think about what else human photographers can capture; some people even began to study Recraft’s color matching and composition, trying to conduct “reverse learning.”

Image source: Xiaohongshu

Although other AI tools have a more delicate painting style and produce more beautiful pictures, they will inevitably have an "AI flavor" after looking at them for a long time. .

Whether it is the disharmonious color transition, the unnatural structure; or the overly smooth and neat outlines, or the flawless texture, in short, what was "fake at first glance" in the early days was AI flavor, but now Too realistic and flawless is also a kind of AI flavor.

In the face of numerous AI painting tools that pursue details and strive for realism, a strong and distinctive visual style is the key to Recraft’s success.

The Hedgehog Commune (ID: ciweigongshe) tried to use several different models and entered the same password "many animals". Judging from the final generated results, Recraft’s built-in Hard Flash mode does have a different feel at first glance.

Recraft does not have a conventional composition like FLUX or DALL·E 3, but arranges different animals in a row. The large areas of land and sky in the picture are left blank, which seems to add a sense of depth to the image. A different emotion: This is a lone lion.

Generated by Recraft, FLUX, and DALL·E 3 from left to right

When the images generated by AI can convey emotions, Recraft AI has also been labeled "Dream Core" With the label "weird", more and more netizens have begun to stimulate their creative desire.

Some people enter their favorite movie lines or lyrics as passwords into Recraft, anticipating what screen will be generated.

In a Xiaohongshu note with 16,000 likes, the blogger "Fan" input lines from "Space Exploration Editorial Department" into Recraft. Although the generated pictures cannot be 100% reproduced The content of the lines, but the overall picture style is simply "more space exploration than the space exploration editorial department."

Some people also try to use Recraft to record their dreams or express indescribable emotions. Some enthusiastic netizens have compiled common prompt words that can make Recraft generate more ethereal and dreamy pictures. Some netizens have even discovered new business opportunities, helping users who cannot use Recraft to generate their ownThe dream core pictures in our hearts cost a few dollars a picture.

Image source: Xiaohongshu

With the enthusiastic participation of netizens, following the Miaoya Camera and Remini clay special effects, another wave of AI carnival has set off. Everyone seems to want to give it a try themselves, input a "spell" into the AI, generate an imaginative picture, and achieve a wonderful feeling of "magic coming true".

As a result, the recent social media platforms such as Xiaohongshu and Douyin seem to be surrounded by "magic": Hello Kitty standing by the window watching fireworks, puppies eating cakes in the snow, goldfish in the blue sky Flying through...

On November 25, Xiaohongshu’s official technology team also launched a special event. Users use Recraft to create pictures and post notes with related topics, and they will have the opportunity to be pushed.

Driven by social media, according to Weidian data, the number of downloads of Recraft in the domestic App Store in the past week has jumped to second place in the "Graphics and Design List".

What is the background of AI dark horse?

Although Recraft AI has captured the hearts of a large number of domestic netizens, in fact, this model neither reflects nor represents the technical strength of Recraft AI.

In the eyes of many users who often use AI drawing tools, many large AI painting models that have been available before can achieve similar film effects through password input.

What really proves the strength of Recraft AI is the Recraft V3 model released this year.

Before the official announcement of the Recraft V3 model, Recraft AI used the pseudonym "red-panda" to participate in voting in the AI ​​image arena on the Artificial Analysis website, and surpassed FLUX, Midjourney, Ideogram, and Stable Diffusion 3.5 in one fell swoop. Ranked first, becoming a dark horse in the AI ​​generated image track.

Because of the name "red-panda", many people initially speculated whether the model was behind a Chinese company. Until October this year, Recraft AI claimed it on Twitter With this model, people have just begun to pay attention to this AI company that has been established for two years.

Recraft AI was founded in 2022 and is a UK-based startup.

Founder and CEO Anna Veronika Dorogush, who previously worked in software engineering at Google and Microsoft, later joined Russia's largest search engine Platform Yandex, creator of the CatBoost open source gradient boosting library.

In January this year, Recraft AI received an $11 million Series A round of financing led by US venture capital firm Khosla Ventures, with former GitHub CEO Nat Friedman also participating. Financing is mainly used to accelerate technology research and development and market expansion.

Since most domestic users learned about Recraft AI through the “Hard Flash” mode spread on social media, people often mistakenly think that it is an AI image generation company like Midjourney.

But after understanding it, you will find that the original intention of Recraft AI has always been to "focus on providing AI auxiliary tools for graphic designers." Therefore, compared to Midjourney, it is actually more like the AI ​​version of Cavan or Photoshop.

Even the No. 1 Recraft V3 model was trained by Recraft AI to a certain extent to facilitate designers to generate posters.

In the official blog introducing the Recraft V3 model, the company claimed that this model is "the only AI image model that can generate long text content in the field of image generation."

For example, if a designer needs to display a large amount of text content on a poster, previous AI graphics models are prone to spelling errors in the text content, so the designer's normal approach may be to first Use AI to generate a background image for the poster, and then use other tools to add the text content.

The logic of the Recraft V3 model is to streamline the process of designers using AI to generate posters by improving the accuracy of AI rendering text content directly in images. The AI ​​media "Xinzhiyuan" once explained the operating mechanism of the model in an article:

In the process of constructing text information, the Recraft team uses the TextDiffuser-2 representation method. Each line of text first records the text's content, and then specify the specific area of ​​text through coordinates. But unlike TextDiffuser-2, Recraft uses three coordinate points to represent text, allowing the model to support rendering skewed text.

In short, the result is that with the help of Recraft V3, designers can greatly improve the generation effect and control of text content in posters.

In addition to AI-generated comic style, realistic style, film style, vector graphics, illustrations, icons and 3D images, Recraft also provides design tools such as lasso, partial redraw, cutout, and mockup.

After experiencing it, what surprised Hedgehog Commune’s design colleagues the most was the mockup function. With the help of AI capabilities, Recraft AI can automatically fit patterns or icons to product images and directly generate product samples, saving designers the trouble of manually adjusting parameters. In addition, ReCraft also introduces a real-time collaboration feature, where designers on the same project can comment on the generated content on an unlimited canvas and make timely modifications.

Understanding the model is not enough, you must also understand the content

Although it seems that there 80% of domestic users do not use Recraft AI as an AI design tool as the founders envisioned, but it is not a bad thing for Hard Flash to become popular.

Judging from the history of the field of AI-generated images, the past three years have definitely been the most intense period of competition among major players in the track. Everyone is scrambling to update large models for fear of being left behind accidentally. behind.

In early 2021, OpenAI released the DALL-E large model, which allows AI to generate images through text;

In March 2022, Midjourney will be launched, which can quickly generate images based on text input by users. High quality images;

August 2022, Stable Diffusion is officially open source and realizes text-to-image generation by converting random noise into high-fidelity images;

In August 2024, Black Forest Labs launched the FLUX model, which improves image quality, text understanding and detail performance. Both surpassed Stable Diffusion 3 and Midjourney and became the new leader in this field...

And in October, the protagonist of people's discussion quickly became Recraft.

According to official data from Recraft AI, 20 days after the release of the Recraft V3 model, the application’s cumulative registered users worldwide have exceeded 2 million. Every time they register and log in, Recraft will ask users “how they heard about the application.” According to founder Dorogush, “Almost all growth comes from social media and word-of-mouth among users. ”

Recraft AI Generation

The implicit change behind this is that as AI-generated image tools gradually enter the lives of the general public, AI companies must rely only on models to figure out how to get around. Strength may not be enough. As Dorogush said in the interview: "It is not enough to generate high-quality AI images. (The product) also needs to build something that can attract people's attention."

In the past. Some AI products that are mainly targeted at the consumer side have actually proven this.

For example, the Miaoya Camera, which became popular on the Internet last year, and the Remini, which became popular in the first half of this year. One of these two products aims at "AI photo taking" and the other creates "clay special effects". Both of them rely on It is the highly representative product functions that leave an irreplaceable product label in the hearts of users.

For RecraftFrom the perspective of AI, the film-like feel brought by Hard Flash is the "handle" for it to break out of the social media circle and form recognition in the minds of more users.

Recraft AI generation

Among the AI ​​models that strive for picture details and realism, Recraft unexpectedly cuts into the track of "imagination" that allows for unreasonableness.

On the one hand, the uniform filter style has deepened ordinary users’ memory of the application and made the name Recraft take root in people’s minds. On the other hand, this is also very clever in covering up some of the deficiencies in the AI ​​model's capabilities. Even if the generated characters are full of plastic and the pictures always have a colorful curtain as the background, under the "Dream Core" and "Weird" style labels Next, everything becomes reasonable.

AI companies continue to pursue breakthroughs in technological capabilities, which is naturally still the top priority in the current AI development stage. But if you want to lead more ordinary users into the AI ​​era, perhaps in addition to breakthrough AI technology, you also need to further lower the threshold for use, choose appropriate product positioning and marketing strategies. Only in this way can more and more users AI products “fly into the homes of ordinary people.”

Reference articles: 1. New Wisdom: In-depth analysis of Recraft V3 to break through text rendering limitations. How did the dark horse of "Wen Sheng Tu" become a dark horse? 2. Web3 Sky City: Why does AI painting advance so rapidly? From history to technological breakthroughs, understand the popular development history of AI painting in one article