ByteDance AI video generation product. Dreaminawas recently appointed as a new mobile leader. Cao Dapeng is the former PopAI head at 01.AI. Cao Dapeng, after joining CapCut, reported to Zhang Nan.
Dreamina also considers using DeepSeek. DeepSeek was already used by Feishu in the ByteDance System.
DeepSeek is a popular tool that has a large user base. This popularity has led to an interactive gameplay. For example, you can use DeepSeek first to create more detailed video scripts, and then create videos using Dreamina. Dreamina users increased rapidly around Chinese New Year. QuestMobile, a third-party monitoring platform, reported that Dreamina had 760 thousand active weekly users at the end December of last year. By mid-February of this year it had nearly 2 million – a three-fold increase in just a month-and-a-half.
Cao Dapeng, Dreamina’s mobile end leader, joined the large-model startup 01.AI in the year 2023 as its product manager for its overseas productivity product PopAI. PopAI was released in August 2023. A year later, 01.AI revealed its user base had reached close to tens millions with a ROI of 1. Cao Dapeng, despite the success of the product, left 01.AI in mid-2024 due to a change in business direction.
Cao Dapeng worked for Feishu, a subsidiary of ByteDance, in 2021. He was the product manager responsible for a specific product. ByteDance is always looking for people with entrepreneurial experience. Cao Dapeng fits that profile. He was involved in the founding of “nice” from 2013 to 2017.
Dreamina, an AI creation platform, is a CapCut product that includes functions such as image generation and intelligent canvas. It was launched in 2024.
Zhu Jun is the head of ByteDance’s main AI department, “Flow”. It includes the dialogue question-and answer product Doubao as well as virtual companion Mao Xiang and AI education product Doubao Ai Xue. ByteDance has a team called “Seed” that is responsible for creating the basic models, supporting ByteDance AI products like Flow.
Dreamina is another important AI product by ByteDance. Dreamina was not under Flow, but under the leadership Zhang Nan at CapCut. Zhang Nan was previously CEO of Douyin Group, and has been responsible for CapCut’s business since early 2024.
SEE OTHERWISE: Zhang Nan Resigns from Douyin Group CEO to Focus on ByteDance CapCut
Sources close to Zhang Nan claim that he has put a lot of effort into Dreamina, and the business is mature. Capcut, which generated revenue of nearly one billion yuan last year, is “not very involved”. Senior executives from ByteDance often meet with Dreamina to share ideas.
ByteDance aims to build the “Douyin”the AI era, within 10 years. However, there is no rush in the short-term. Dreamina was always a priority for ByteDance but it was initially limited by its insufficient model capabilities. As a result, user data comparisons with products of the same period were not high. In December last, the video creation product PixVerse, from startup Aishi Technology, had almost ten million active monthly users, whereas Dreamin
Zhang Nan’s OKR 2025 includes a focus to refine Dreamina’s effectiveness.
ByteDance Seed Model Team also created the video models and speech model used in Dreamina. The Seed team includes talents such as Yang Jianzhao Feng Jiashi Wang Yuxuan Tian Zhi Jiang Lu and others who are researching and developing models. The research and development team of the Pixel Dance video model, which was developed by AI Lab in the second half last year, was also transferred to Seed.
ByteDance places a high value on model research and development. In January of this year, they established AGI Research Program – Seed Edge. In February, they recruited top talent Wu Yonghui to join Seed.
SEE ALSO: ByteDance Adjusts AI Department “Seed”Yonghui Wu Becomes New Head
The industry’s current view is that model intelligence and products are not directly linked because the data from user-product interaction and dialogue are not of high quality. Instead, the improvement in model intelligence relies on R&D team’s judgements about technological trends, rather than user feedback.
Before DeepSeek’s multi-turn conversational product was launched on January 15, this year, users had no direct way to interact with the company. The model was still impressively intelligent.
ByteDance described an improvement cycle that involves model feedback and improving the experience of users. Byte’s official blog stated that in January of this year, when Byte released its Douyin 1.5 Pro model with multimodal capabilities and integrated into their official blog:
Byte, based on extensive user feedback from Doubao, has built a closed-loop optimization system that includes problem discovery, data mining, human-machine combined annotation, rapid iteration through a flywheel of user data, continuously enhancing the actual usage experience of models. Doubao has provided extensive user feedback that they have used to build a closed loop optimization system. From problem discovery, data mining, human-machine combined annotations to rapid iteration using a flywheel with user data continually enhancing the actual use experience of models.
ByteDance described one aspect of the flywheel as “actual usage experience of model” rather than “model performance”. This is not in conflict with the current mainstream understanding, but it also shows ByteDance’s desire to find an’snowball’ that can roll up into large-scale modeling business: after all, if such flywheels exist they could be good news to companies that have efficient systems for collecting feedback data from users and control significant numbers.
This also indicates possible organizational changes: models should be under modeling, while (product experiences) should be under (product experiences). Sign up for 5 articles free per month !
