Lovart Founder Chen Mian: From a Cash Crisis to a Design Agent

Zhang Xiaojun introduces Lovart founder Chen Mian, whose product had recently attracted attention when interviewed in 2025. She asks about two years of subsidy competition, a product suspension, funding difficulties and the search for a new direction.

Chen says the parent company was founded in May 2023 and the Lovart operating company later. The business was not profitable at the time; he declined to give a precise valuation, saying only that it was in the hundreds of millions of dollars.

During university he interned at Tencent and 360 on products including Qzone and a mobile app assistant, learning to focus on user needs, solve a specific problem and refine the experience.

After graduation he worked at 360, Didi, Mobike, Missfresh and ByteDance. Competition at Didi and Mobike showed him both subsidy-driven growth and the business-model constraints that followed.

Using bike sharing as an example, he says orders can grow rapidly when fleet and maintenance costs are ignored; once a company needs to break even, fleet size and user prices can change and the market must be recalculated.

In his assessment, Mobike had limited profit and traffic potential as a standalone business. Helping a payment strategy did not automatically prove its own business model; he frames this as a lesson from his experience.

He recalls how enthusiasm during rapid growth can obscure doubts about costs. After Mobike was acquired and began examining economics more closely, he says the issues became clearer.

At Missfresh he led consumer product, operations and growth. Despite growth, he concluded that low retail margins and an offline supply chain limited the improvement internet tools could provide.

The experience made him prioritize business models and return to online products. He describes retail as demanding disciplined execution and creative products as requiring empathy and imagination, a judgment about his own fit.

At ByteDance he worked on the education product Guagualong and later moved to Jianying, handling its China business and commercialization for Jianying and CapCut globally. Frequent moves exposed him to many businesses’ rises and setbacks.

He once regretted not founding a company earlier and worried he had missed the mobile-internet window. The emergence of AI made him think his varied experience could be useful in a new venture.

He believes online, content and creative work suit him better and chooses directions according to the team’s strengths. At the time, he did not plan a near-term enterprise-services push.

He praises the pragmatic debate, fast decisions and product quality he observed at ByteDance. Earlier jobs helped him recognize mistakes; ByteDance also let him watch how others made effective decisions.

One approach is to envision an industry’s eventual shape and work backward to rule out dead ends. He stresses that this cannot replace practical attention to each immediate product decision.

He attempted a GitHub-like project in university but chose to learn at larger companies. By 2021, successful projects and discussions with colleagues had strengthened his confidence to start a business.

As ChatGPT, Stable Diffusion and Midjourney entered public view in late 2022, he judged that AI could have a deeper impact than mobile internet and decided to start a company. This was his assessment at the time.

He sees mobile internet as an expansion of computing and internet use cases, while viewing AI as a more fundamental change in machine intelligence. He argues that the previous product cycle alone cannot predict the next one.

From the outset he chose applications rather than foundation models: his background was product management rather than model research. He wanted to build differentiation outside the base model.

He separates information production and consumption into text, images, video, search, social and entertainment. Because multimodal creation did not fully overlap with a general model interface, he chose creative work as the entry point.

He sees differentiation in domain data, assets and interaction. The first product used assets and templates; Lovart combines conversation, an editable canvas and tools.

Chen says the company’s original focus was creation, with design as one part. He predicts AI will change both creative tools and the division of work between people and those tools.

Early on he asked a technical partner to validate the product idea and then met investors. He says the first round raised about $4 million at an approximately $18 million valuation, and later considered not raising more promptly a mistake.

The team soon faced image-generation competition. He says it subsidized creators and users in 2023 and spent several million dollars across subsidies, staffing and GPU costs.

In September 2023 the first product was taken down over a required filing. Chen says the team had not understood the requirement beforehand; cash was dwindling and the company quickly cut staff.

He says the team grew to roughly 50 or 60 at the height of competition and then fell to about 30. His claim of leadership in China’s image-generation market is his own, not an independently verified ranking.

He still believes competition was necessary but admits neglecting cash management and fundraising timing as a first-time founder. He cut spending and pursued paid memberships while waiting for the product to return.

By March 2024, he says, the company’s account briefly held only about RMB 4,000 before a crucial investment arrived. The smaller team included people taking reduced or deferred pay, and he was prepared to contribute his own money.

On user acquisition, he emphasizes innovation and timing: when a product already attracts interest, advertising can amplify it. He puts the first product’s early acquisition cost at roughly RMB 3–4 per user, with competition potentially raising costs.

During the suspension he met multiple groups of investors daily, many concerned about whether the product and users would return. He recalls frustration with repeated conversations ending in rejection and learned to prepare during good periods for bad ones.

He says the filing, monetization and user growth improved in 2024, followed by several funding rounds. With cash tight, he sought staged closings rather than waiting to combine every investor into one round.

Although powerful, the first product had a high learning curve even for professional designers. The team wanted the next product to preserve AI capability while making it easier for more people to use.

His longer-term vision is AI as not only a designer’s tool but potentially an assistant in creative work. He says this depends on progress in both model capability and product interaction.

The team began preparing canvas capabilities in the second half of 2024. It first imagined users building workflows on the canvas, but found node-based construction unnatural for its target designers.

As model agent capabilities improved, the team tried having AI plan the workflow. Chen compares the interface to a desk, toolbox and conversation: users indicate changes and the system calls tools.

He estimates three to four months from the first idea to the product and says the emergence of products such as Manus increased the urgency. He sees Cursor as an early specialized agent and wants Lovart to focus on creation.

The team chose Lovart as a short, pronounceable name with an artistic feel. He recalls registering only a few domains and finding other suffixes taken after launch.

He says early attention and brand help overcome a cold start, but long-term performance depends on the product. At the interview he reported better retention than the prior product and about 120,000 people on the waitlist.

Invitations reflected early experience concerns, unfinished payment features and the cost of free usage. He says the team intended a test and did not expect so much attention.

He recalls rapid growth in social-media views and sharing after launch. Feedback in design and AI communities helped with the cold start, though he acknowledges that this was attention within relevant circles rather than universal adoption.

To meet the launch window, the team worked at high intensity. He says he often stopped around 3 a.m. and explicitly calls the schedule unhealthy.

He considers domain knowledge and teaching it to AI important, while acknowledging the team is just beginning and must still prove long-term retention and paid demand.

Looking ahead, he expects general and specialized agents to coexist for a period: general entry points handle broad needs, while specialized tools provide depth and may also serve users directly.

If intelligence advances far beyond current levels, the structure could change again, but he considers that premature. His claims about market concentration and technical progress are forecasts.

He predicts a small number of leaders in design agents alongside products from major ecosystems. Subscriptions may fit professionals using a tool continuously; per-output pricing may fit customers buying a single result.

At the interview he was focused on North American growth, collaboration with the China team, product iteration and technical challenges. He expected larger companies might enter and saw early advantage and user understanding as important.

He describes simplifying product development around domain experts and engineers who can program with AI, while retaining functions such as marketing. This is his observation about his company, not a prescription for every team.

He says the company then had nearly 100 people, with substantial operations for the first product and a smaller team on the second. He considers product and strategy his strengths and describes his style as more emotional.

He admits being demanding and sometimes harsh at work and recognizes that emotional swings can tire the team. His ambition for the company’s size is a goal, not an achieved result.

He says a startup can fail through poor cash management, fundraising timing, excessive expansion or inadequate preparation for competition. Before Lovart gained attention, he worried others might build the next product first.

His greatest excitement is seeing a product hypothesis receive user feedback. He says he read many designer responses and saw encouraging signals for Lovart, while recognizing that substantial challenges remained.

Some founding team members left during the difficult period, which disappointed him; in retrospect he understands their choices and values those who stayed.

He compares innovation to repeatedly striking matches against damp wood: attempts can fail many times, and timing, patience and learning matter. He says he cannot know whether he will ultimately be the one to succeed.

He moved frequently as a child as his family and schooling changed cities. Computer games sparked his interest in computing, while Jin Yong’s martial-arts novels shaped his wish to create something meaningful.

For him, personal success means creating something that makes the world a little different and preferably better; he adds that others can define success differently.

He advises other founders to focus on specialized use cases, domain data and knowledge and prepare for model improvements. His belief that general models cannot immediately cover every specialized need is a product and investment hypothesis.

He says participating in a technological wave matters and acknowledges that starting earlier might have taught him more. He connects team cohesion to shared hardship and positive feedback.

He recommends Jin Yong’s novels, a Chinese children’s encyclopedia and Records of the Grand Historian, and mentions Crossing the Chasm. He says the former shaped his values and worldview while the latter helped him understand differences between early and later users.

A key hypothesis is that general models will not quickly replace every specialized application, with a further assumption that AGI will not arrive within five years. He expresses concern about AI broadly surpassing humans and calls safety and boundaries shared questions; these are his views at the time.

Zhang closes the interview.