ByteDance’s Dreamina Is Considering Integrating DeepSeek

ByteDance’s AI video generation product, Dreamina, recently appointed a new mobile leader. He is Cao Dapeng, the former head of PopAI at 01.AI. After joining, Cao Dapeng reported to Zhang Nan, the head of CapCut.

At the same time, Dreamina is also considering using DeepSeek. Previously, Feishu within the ByteDance system had already used DeepSeek.

The popularity of DeepSeek has led to overflow demand and interactive gameplay: for example, first use DeepSeek to generate more detailed video scripts and then create videos in Dreamina. Around Chinese New Year, the number of Dreamina users grew rapidly. According to third-party monitoring platform QuestMobile, Dreamina had around 760 thousand weekly active users at the end of December last year and nearly 2 million by mid-February this year – an increase close to three times in a month and a half.

Dreamina’s new mobile end leader Cao Dapeng joined the large model startup 01.AI in 2023 and served as the product manager of its overseas productivity product PopAI. PopAI was launched in August 2023, and a year later 01.AI revealed that its user base was close to tens of millions, with an ROI close to 1. Despite the good performance of the product, in mid-2024, Cao Dapeng left 01.AI due to its business direction adjustment.

Before joining 01.AI, Cao Dapeng joined Feishu under ByteDance in 2021 as the product manager for a certain product. ByteDance has always liked recruiting people with entrepreneurial experience, and Cao Dapeng fits this profile. From 2013 to 2017, he participated in founding the image social community “nice”.

Dreamina is an AI creation platform under CapCut that supports functions such as image generation, intelligent canvas, and video generation. It was launched in May 2024.

The main AI department within ByteDance is “Flow”, led by Zhu Jun. It includes dialogue question-and-answer product Doubao, virtual companion product Mao Xiang, and AI education product Doubao Ai Xue. The team responsible for developing the basic models at ByteDance is “Seed”, supporting ByteDance’s AI products such as Flow.

Another important AI product from ByteDance is Dreamina. Dreamina is not under Flow but has been under the leadership of Zhang Nan in CapCut. Zhang Nan previously served as CEO of Douyin Group and has been in charge of CapCut business since early 2024.

SEE ALSO: Zhang Nan Resigns as CEO of Douyin Group to Focus on ByteDance’s CapCut

According to sources close to Zhang Nan, he has devoted a lot of energy to Dreamina and the business is quite mature. Capcut, which generated nearly one billion yuan in revenue last year, “is not very involved.” “Dreamina is personally led by Zhang Nan and involves deep daily decision-making; while Capcut is mostly managed by Zhang Nan -1.” Senior executives at ByteDance also frequently meet with the Dreamina team to share ideas.

Building the “Douyin” of the AI era is ByteDance’s goal within 10 years, but there is no rush to be aggressive in the short term; Dreamina has always been a high priority within ByteDance, but early on it was limited by insufficient model capabilities, so user data comparison with products from the same period was not high: In December last year, the video generation product PixVerse from startup Aishi Technology had nearly ten million monthly active users, while Dreamina only had around one million.

Zhang Nan’s OKR for 2025 includes a focus on refining Dreamina’s model effectiveness.

The video models and speech models used behind Dreamina are also achievements of ByteDance’s Seed Model Team. Within the Seed team, talents researching and developing related models include Yang Jianzhao, Feng Jiashi, Wang Yuxuan, Tian Zhi, Jiang Lu etc. In the second half of last year, the research and development team for Pixel Dance video model from AI Lab was also transferred to Seed.

ByteDance is increasingly valuing model research and development. In January this year they established AGI Research Program – Seed Edge; in February they poached top talent Wu Yonghui from Google to join Seed.

SEE ALSO: ByteDance Adjusts AI Department “Seed”, Yonghui Wu Becomes New Head

The mainstream view in the industry now is that model capabilities and products are not directly related, because user-product dialogue and interaction data are not high-quality data, making it difficult to improve model intelligence; the current improvement in model capabilities relies on the R&D team’s judgment of technological trends rather than user feedback.

Before DeepSeek launched its multi-turn conversation product on January 15 this year, there was no direct channel for interacting with users. However, the model still demonstrated impressive intelligence.

ByteDance described a positive cycle of improving model experience and user feedback. In January this year when Byte released the Douyin 1.5 Pro model integrating multimodal capabilities, their official blog stated:

“Relying on AB Test experience in recommendation, search, and advertising fields, Byte has developed an efficient PostTraining full-process based on user feedback. Based on extensive user feedback from Doubao, they have built a closed-loop optimization system from problem discovery to data mining to human-machine combined annotation to rapid iteration through a flywheel of user data continuously enhancing the actual usage experience of models.”

ByteDance used “actual usage experience of models” instead of “model performance” to describe one aspect within the flywheel. This does not conflict with current mainstream understanding but also shows ByteDance’s desire to find a ‘snowball’ that can roll up in large-scale modeling businesses: after all if such a flywheel truly exists it could be good news for companies that have accumulated efficient systems for collecting user feedback data and control significant numbers of users and traffic entries.

It also indicates possible organizational adjustments: models belong under modeling while (product) experiences belong under (product) experiences.