Chen Mian on Two Years of AI Design Entrepreneurship, from a Cash Crisis to Lovart

Zhang Xiaojun interviews Lovart founder Chen Mian about his mobile-internet career, early AI startup competition and funding pressure, and his experience after launching the design agent Lovart. The conversation captures a founder’s product and industry views in 2025.

Chen says the company was founded in 2023 and Lovart’s operating entity was established later. At the interview he says it was not profitable; his funding, valuation, and operating descriptions are self-reported and time-specific.

During university he interned at Tencent and 360 as the industry shifted from PCs to mobile internet. He says focused execution, product detail, and attention to users were foundational lessons.

After graduation he worked on businesses including Didi, Mobike, and Missfresh. He describes Didi’s competition with Kuaidi and Uber and Mobike’s rivalry with ofo as two battles: peak excitement did not necessarily mean the business model was sound.

Using bike sharing as an example, he argues that subsidies and abundant bikes can produce large order counts, but scale is not healthy profit if supply costs and low transaction values are ignored. Reexamining the economics after an acquisition sharpened his understanding.

At Missfresh, he saw that online growth could not easily remove the low margins and offline complexity of retail supply chains. That lesson led him to favor online businesses better suited to his skills.

At ByteDance he helped build the education product Guagualong from scratch; after policy changes affected the business, he moved to Jianying and worked on domestic product and CapCut’s global monetization. Frequent job changes gave him broad experience but also caused doubts about his direction.

He reflects that starting a company earlier or joining a continuously growing business could have been another path. Once the AI wave arrived, he saw his experience of success and failure across companies as useful for choosing a startup direction.

Chen believes online creative products requiring empathy, imagination, and product innovation suit him better than offline retail organizations emphasizing execution and discipline. He frames this as a judgment about himself, not a rule for every industry.

On ByteDance, he says he learned to focus discussions on problems and make decisions quickly, while observing how successful businesses were built. Knowing why something failed and knowing how to build something that works are different experiences.

He says a team can consider a market’s possible long-term shape to rule out paths unlikely to work. That does not replace practical attention to users and product details in the present.

He briefly tried starting a business at university, then went to large companies because he lacked confidence. By 2021 he felt more capable of founding a company, but worried the best mobile-internet window had passed.

The arrival of ChatGPT, Stable Diffusion, and Midjourney convinced him to enter AI. He believes AI could bring a more fundamental change than mobile internet, giving his varied past experience a new use.

He remembers telling his wife that he wanted to start a company and receiving her support. For him, hope gave past setbacks a new meaning and encouraged him to try again.

From the start, the team chose applications over training a foundation model. Chen says his background is product rather than model research, and he wanted multimodal creative use cases less directly overlapping with general models.

He separates information production from consumption. Documents and search are close to language models’ core capabilities, while images and design may offer different data, interfaces, and user needs, making them a starting point for the team.

The first product, LiblibAI, drew on creator assets and templates; the second, Lovart, combines a canvas, conversation, and editing. Chen says differentiation can come from domain data and interaction design as well as models.

He sees design as part of a broader creative field and expects AI to change both creative tools and the division of work between people and tools. Images matured earlier, so the team started there.

At the start, he asked a technical partner to test the product idea before speaking to investors. The first financing was relatively smooth, he says, but the team did not prepare promptly for later capital needs.

Competition in image generation intensified quickly in 2023. The team subsidized creators and users and paid GPU costs for free usage. Chen believed they needed to outperform rivals in efficiency and strategy.

He recalls that the team focused on competition and did not complete the relevant filing in time, leading to the product being taken down in September 2023. Layoffs, insufficient revenue, and fundraising difficulties followed, in his account.

Zhang asks whether a subsidy battle was worthwhile when the product was not in its final form. Chen still believes they needed to win users, but acknowledges inexperience with cash flow and fundraising timing.

He identifies the combined cost of staff, compute, and subsidies as a source of pressure. The takedown, uncertainty about users returning, departures, and investor concerns all made fundraising harder.

After shrinking, the team pursued paid memberships. Chen says the company had about 4,000 yuan left in March 2024 before receiving a critical investment; he had also prepared to use personal funds if needed.

On acquisition, he says advertising is only a channel: product novelty, public attention, and timing all affect cost. A startup cannot assume that simply buying traffic will sustain growth indefinitely.

He recalls meeting multiple investors each day during the difficult period and repeatedly being turned down. They worried whether the product could return; he learned to prepare capital and time for future difficulties while conditions are favorable.

In 2024, after the product returned and revenue and users recovered, the company arranged several funding rounds. Chen reports the amount and timing as his recollection from that period.

He says the first product was powerful for professionals but difficult to learn, attracting designers willing to invest in AI skills. The next generation needed to make capabilities accessible to more users with less friction.

The team began preparing a canvas in the second half of 2024. Initially it considered letting users connect workflow nodes, but found that approach did not fit some designers’ habits, so it kept looking for a more natural interface.

By year’s end, improved agent capabilities suggested letting AI plan workflows for users: the canvas as a work surface, chat for communication, and tools and an editor for execution and revision. The team refined interactions around those components.

He says the journey from idea to product took several months, and Manus’s appearance increased the urgency to ship an agent. Early attention helps, he argues, but lasting performance depends on product quality.

The name Lovart was intended to sound like a designer who loves art and be usable across languages. He jokes that the team registered only a few domains and others were taken after launch, a lesson in penny-pinching.

At the interview the product was invite-only. Chen says payment features were still being built and free usage incurred compute costs, so the team needed to control access and planned to expand it as monetization became ready.

He recalls social-media attention rising rapidly after launch, with discussion in design and AI circles. The initial exposure helped with cold start, but he acknowledges the team still needed to prove long-term retention and willingness to pay.

The team worked intensely to meet its launch window. Chen admits he often worked late into the night and that the pace was unhealthy, hoping it could improve afterward.

In product organization, he sees domain expertise and the ability to convey it to AI as increasingly important. He also stresses that the team was only beginning with agents and still faced the tests of sustained use and monetization.

He predicts general and vertical agents could coexist: general entry points handle broad needs, specialized tools handle deeper tasks, and users may also visit vertical products directly. This is his industry hypothesis.

On model companies and application companies, he says providers may offer features close to base-model capabilities, while some vertical applications also need distinct data, interfaces, and domain knowledge.

He expects leaders in design tools alongside products from large companies within their ecosystems, though the competitive outcome remains uncertain. In the near term, his team was focused on professional designers.

The business model also depends on the user. Professionals who use a tool continuously may suit subscriptions; if a product eventually delivers a design directly to a client, Chen imagines charging by output type and value.

At the time he was focused on North American growth, cross-region teamwork, and product iteration, while preparing for large companies that might enter creative tools. Early action, user recognition, and continued innovation all matter, he says, but none guarantees victory.

Chen describes changing team roles: domain experts, engineers who code with AI, and model researchers need closer collaboration. He presents this structure as his company’s experiment, not a template for every team.

He acknowledges anxiety in founding a company, especially when the first product has limitations and the next one may not succeed. He chooses to work through that pressure while recognizing further competition ahead.

His happiest moments come when user feedback supports an idea about a future product. He says Lovart’s early attention is only a beginning; what excites him is bringing an imagined product into reality.

Compared with a large-company rank, he now values the work created and personal growth more. Titles and pay serve management purposes, he says, but cannot fully measure what a person has accomplished.

At the hardest point, some core team members left, disappointing him. Later he understood that people make different choices under risk and came to value those who stayed even more.

He compares innovation in an unfamiliar field to repeatedly trying to light damp wood: there are failures and brief sparks, and patience and timing matter. He cannot be certain he will be the one to succeed, but wants to keep participating.

He recalls a moment in an investor’s office while fundraising in late 2023: he realized he was living the founder’s path he once admired. Even without knowing the outcome, the thought renewed his energy.

Frequent moves in childhood made it difficult for him to identify a hometown. Computer games sparked his interest in technology, while Jin Yong’s fiction shaped his idea of doing meaningful work; he hopes entrepreneurship connects those experiences.

His advice to other founders is to pay attention to specific industries, data, and needs while watching model capabilities evolve. He believes vertical applications have room, an outlook that depends on his assumptions about the pace of general intelligence.

He recommends Jin Yong’s fiction, a children’s encyclopedia, and Records of the Grand Historian as influences on his values and worldview. He also expresses unease about AI exceeding humans broadly and says safety and boundaries deserve shared discussion.

Zhang closes the interview and invites listeners to share topics for future episodes.