Our Pivot Story

We launched our first product ChatDesigner on Product Hunt on November 17 last year, where it was voted #3 product of the day. This August 17th, we officially shut down chatdesigner.ai, and we are now launching HuHu.ai, a new virtual try-on product for clothing eCommerce sellers this Sunday (September 1). Here’s a story of what happened over the past 9 months. ChatDesigner was built with the hope of marrying LLM and image editing, allowing users to directly chat to edit their images. The concept sounds cool, which is why we were able to grow from 0 to 70K users purely organically. But the reality is cruel. To achieve the goal, we had to develop and maintain over 40 different AI features in the product. Despite our efforts, the limitations of AI and LLM meant that we still couldn’t fully cover all the use cases users wanted. Many users signed up, tried out, felt disappointed and left. In the meantime, the competition in the market is fierce, with many players offering daily refreshed free credits, we were unable to raise our price. Back in February, during the third month of Chatdesigner’s launch, our revenue was low and the user growth was stagnant. We ran a series of funnel tests, but the average LTV of Chatdesigner was far below the customer acquisition cost (mainly GPU cost for offering free credits). In May, we interviewed with Y-Combinator. The interviewers asked us bluntly: “How much longer do you need to realize your business model doesn’t work?” We were tongue tied, but at the same time realized we had to pivot. The decision to pivot was easy, but deciding which industry to pivot to and how to do it was much more difficult. At that moment, some corporate clients approached us with the needs for AI virtual try-on. Although it seemed like a space with many players, the industry was still not satisfied with the results. We realized this could be our opportunity. Virtual try-on is not a new demand; it’s been around for over 10 years. From purely 3D approaches to today’s diffusion-based methods, new players have emerged continuously, but none have become dominant. The fundamental issue remains that the technology has not reached the threshold for large-scale commercial use. When we began researching this field, we had one question in mind: has the scaling law in this area finally been established? If the answer is yes, then we are at a turning point in this technology. We just need to outperform the state-of-the-art (SOTA) solution slightly to gain a first-mover advantage; otherwise, we would fade into the background. However, we were afraid to put all our eggs in one basket, as we weren’t 100% sure we could surpass the SOTA and achieve commercialization. We split into two small teams: one focused on training the virtual try-on model, and the other explored another B2B direction. Fortunately, the results were promising. In early July, we released the v1 model internally, and user feedback indicated that it had already surpassed SOTA, but SOTA doesn’t necessarily mean it’s ready for large-scale commercial use. Our v2 model, released in August, further strengthened common use cases apparel sellers need in the market. When we tested the model internally, we were often surprised by its capability. And the results have already been well received by many local sellers during our demos, even though it wasn't publicly available yet. We know that now it’s time to launch the new product. After going through our darkest hours, we have a few thoughts to share: First, know who you are before choosing your entrepreneurial direction. For a team led by mostly engineers and scientists, it’s best to choose a direction where the business case is clear but the technology has yet to break through. In our scenario, there will undoubtedly be many so-called competitors in the market, but don’t panic. The key factor in success is whether you can show your customers that your solution is significantly better than others they’ve seen. Our biggest mistake early on was pursuing a growth and operations-driven direction as a tech-focused team. Second, doing one thing better than anyone else is more valuable than being mediocre at a hundred things. If you need to multiply this by a factor, it should be a thousand times better. During the chatdesigner period, we developed over 40 features in one month, but none were impressive, and the team was overwhelmed. And since the product is not significantly better than others, it’s also difficult for us to raise our prices. Third, this is an era where experience is becoming less valuable. The history of Stable Diffusion and LLMs is not long, especially Stable Diffusion, which was invented in August 2022. In this era, what matters are intuition, curiosity, and energy to experiment. Experience itself may even become a burden when it comes to taking bold actions. This is an age where everyone is mining. People are digging with their shovels, not knowing what lies beneath the ore. If we strike gold, it will ultimately belong to all of humanity. The reward we receive is the ticket money for the privilege of discovering the treasure, but we don’t own any treasure. The most interesting part of the whole process isn’t the reward itself but the expectation of what treasure lies beneath the ore. If you are interested in our AI virtual try-on product, please visit our website https://www.huhu.ai. We’ll launch on Product Hunt this Sunday (Sep 1) with free credits to try out. Looking forward to your support: https://www.producthunt.com/products/huhu-ai