Benefiting from the Manus dividend, this overseas design Agent has 20,000 people queuing up to apply! | Emergence of NewThings

Text by | Zhou Xinyu

Edited by | Su Jianxun

Lovart is definitely one of the first products to reap the Agent dividends brought by Manus.

After its release on May 13, 2025, this AI design application, touted as "the world's first design Agent," had 20,000 people queuing up to apply on the day of its launch, and the invitation codes were once resold for 500 yuan each. Tech influencers like Elon Musk and Santiago also liked or commented on it on X.

Before the hype around Lovart subsided, on July 3, 2025, its local version, "Xingliu Agent," was launched. This is also the first Agent application to be locally adapted after the release of Manus. In contrast, the Chinese version of Manus, which was launched in March, is still under development.

△ Xingliu Agent. Image source: Xingliu AI official website

Behind it is LiblibAI, founded in May 2023. It is currently the AI image startup in China with the highest total financing amount and the image model creation platform with the largest number of creators (over 20 million). Its founder and CEO, Chen Mian, was once the global commercialization head of ByteDance's Jianying and CapCut.

In a media interview, Chen Mian described the current characteristics of AI startups as "leading in cognition + extreme execution": For innovative companies, being able to create the first innovative product is very important.

Currently, Lovart and Xingliu Agent have respectively filled the gaps in the design Agent ecosystems overseas and in China.

However, due to the different models they are connected to, local adaptation for Agents inevitably results in some performance and experience being sacrificed. To live up to Lovart's reputation, Xingliu Agent still has some way to go.

Xingliu Agent is localized, but I still choose Lovart

As a novice design user, after using both Xingliu Agent and Lovart, the author's biggest impression is:

Both Agents have good aesthetics, but in terms of text understanding, Lovart still has an edge over Xingliu Agent.

During the use, the author input the same prompt into both Lovart and Xingliu Agent, which included displaying TikTok elements in the picture and a piece of code.

The generated results showed that in the two pictures provided by Xingliu Agent, the word "TikTok" was misspelled, and the provided code was a jumble. As for Lovart's code, regardless of whether it could run, its spelling and format were almost correct.

△ A Pop art illustration with elements like TikTok and code generated by Xingliu Agent. Image source: Author's test

△ A Pop art illustration with elements like TikTok and code generated by Lovart. Image source: Author's test

The root cause lies in the models they are connected to.

Behind Lovart, there are a variety of overseas SOTA models, including GPT image-1, Flux Pro, OpenAI-o3, and Gemini Imagen 3.

For example, GPT image-1 is currently OpenAI's most powerful multimodal model, with outstanding text rendering capabilities. In simple terms, it can more reliably generate accurate text in images to meet the needs of picture design with text descriptions.

The operation panel shows that during the image generation process, Xingliu Agent mainly calls models like Flux.1. Although Flux performs well in terms of consistency, there is still a gap in text rendering capabilities compared to models like GPT and Gemini.

The official claims that Xingliu Agent is more suitable for Chinese users and has a better understanding of Chinese semantics. However, in actual operation, there isn't much difference between Lovart and Xingliu Agent in understanding Chinese aesthetics and elements.

Comments.

Taking the generation of a poster for the TV series "Empresses in the Palace" as an example, both Xingliu Agent and Lovart produced good results, with the posters featuring Qing Dynasty costumes and palace elements. In fact, the author personally believes that Lovart has better aesthetics in some designs.

However, a major flaw is that both Xingliu Agent and Lovart failed in text rendering on the posters. Xingliu Agent misspelled one character in "Empresses in the Palace," while Lovart misspelled two.

△ A poster for "Empresses in the Palace" generated by Xingliu Agent. Image source: Author's test

△ A poster for "Empresses in the Palace" generated by Lovart. Image source: Author's test

In other functions such as image editing and video generation, there isn't much difference between Xingliu Agent and Lovart. For example, the author used Xingliu Agent to search the Internet and generate a set of new visual images for "Intelligent Emergence." The author could also directly edit the images through text conversations.

△ Editing an image through conversation. The right picture shows the edited effect. Image source: Author's test

Like Lovart, Xingliu Agent is also connected to video and 3D models such as Kling 2.1 and Hailuo 02. Therefore, Xingliu Agent can also smoothly generate videos with BGM from image storyboards.

The author used Xingliu Agent to create an animated effect for the "Intelligent Emergence" logo with BGM. From generating the animation storyboard, to selecting the background music, to video generation, the Agent only needed one round of conversation, and the experience was quite good.

△ An animated logo for "Intelligent Emergence" created by Xingliu Agent. Image source: Author's test

Moreover, relying on the rich creator and model ecosystem that LiblibAI has built in China, Xingliu AI has a stronger "community feel." For example, the built - in "Image Inspiration" interface in Xingliu AI is not available in Lovart yet.

The rich categories such as "Art Illustrations" and "Cartoons" provided there can better inspire creators. For novice users, they can also create images in the same style with one click.

△ "Image Inspiration" provided by Xingliu AI. Image source: Xingliu AI official website

△ Lovart only has works generated by the Agent, without detailed categories. Image source: Lovart official website

However, this Agent, which claims to be for professional designers, still has a long way to go in the real design and creation field.

For example, many designers have reported on social platforms that currently, the Agent mainly focuses on "one - click image generation" based on the model's understanding ability. It lacks data accumulation in professional scenarios and is difficult to adapt to workflows with fixed SOPs (Standard Operating Procedures).

"Currently, design Agents are more like small tools to meet the occasional creative needs of ordinary people," a graphic design professional told us. "In the professional field, these Agents can only assist in some aspects and cannot meet all the needs 'in one stop'."

Quickly replicate proven products and experiences to untapped markets

LiblibAI's two most successful products, the image model community LiblibAI and the overseas design Agent Lovart, both follow the same strategy: Replicate proven products and experiences to an untapped market to gain a first - mover advantage.

For example, when LiblibAI was founded, it targeted Civitai, an overseas image model community with over 3 million registered users. Coincidentally, the domestic image model community market was still unexplored at that time.

Lovart's rapid growth overseas is also due to replicating Manus' approach in the design vertical, an area where Agents had not yet entered.

An employee of LiblibAI revealed to "Intelligent Emergence" that LiblibAI started planning for Agents in early 2025 but couldn't determine the product form until Manus was released - Three days later, the Lovart team was immediately assembled. In the end, both the product form of "task creation area + process visualization" and the marketing strategy of issuing invitation codes of Lovart were similar to Manus.

Speed is another key factor for LiblibAI to establish early awareness among users.

This has been ingrained in LiblibAI since its founding. Zhang Zijie, the co - founder of LiblibAI, once told "Intelligent Emergence" that if LiblibAI had launched one month later, there would have been two or three more competitors in the market, and subsequent financing and user growth would have been much more difficult.

Many employees recalled that in order to seize the first wave of dividends from Manus, Lovart's R & D schedule was quite aggressive: The company initially aimed to complete the R & D within three weeks, but later changed it to one month for feasibility reasons. Employees worked six - day weeks, and working overtime until after 10 p.m. was common.

Facts have proven that Lovart, which launched early, did continue Manus' popularity: Within five days of its release, more than 100,000 users flocked to its Discord channel. According to data from Liangziwei Think Tank, in the month of its release, Lovart's total monthly web visits reached about 300,000.

Xingliu Agent is now facing a more competitive domestic market. Tech giants like ByteDance are aggressively squeezing the space of domestic image startups.

According to Feifan Research data, from March to April 2025, Xingliu AI's monthly visits dropped from 1 million to about 500,000. In contrast, ByteDance's Jimeng AI still maintained a growth rate of 5.92%, with monthly visits reaching about 6.75 million in April.

After building momentum overseas, quickly localizing, it's obvious that Xingliu Agent aims to inherit Lovart's traffic. An employee of LiblibAI also said that among Lovart's users, half are Chinese, "indicating that there is indeed a demand among Chinese people."

Whether Xingliu Agent can drive the growth of Xingliu AI by leveraging Lovart's name remains to be tested by the domestic market.

Welcome to communicate!