Two former ByteDance executives go solo to challenge their old employer?

ByteDance's AI video generation product, Dreamina, has long been a front-runner in the industry. Its leading position stems not only from the refinement of its large model, but also from ByteDance's powerful built-in distribution channels. After integrating with Doubao and CapCut successively, its traffic has consistently ranked among the top in the industry, and the launch of Seedance 2.0 has ushered in a new milestone for its traffic growth.

Right when Dreamina is occupying most of the market share, two rising star products founded by former ByteDance executives are drawing widespread attention.

Launched in June 2025, "Paiwo AI" is a product of PixVerse Tech. Its founder and CEO Wang Changhu was the former head of AI Lab Vision at ByteDance, who built the visual technology system for Douyin and TikTok from scratch. PixVerse Tech announced the completion of its Series C financing in July 2026, with total funding reaching 2.98 billion RMB, backed by investors including Alibaba (leading multiple rounds), CDH Investments, China Ruyi, and 37 Interactive Entertainment.

LiblibTV, launched in March 2026, is a product of Eanyo Tech. Its founder Chen Mian was once the global commercialization lead for Jianying and CapCut at ByteDance, and also the youngest Level 4-1 executive within ByteDance. Eanyo Tech officially announced in June 2026 that it had completed nearly $300 million in Series B+ financing, with a valuation exceeding $2 billion, setting a new financing record for the domestic AI application track. This round of financing was jointly led by GraniteAsia, Tencent, and Shunwei Capital, with participation from HTInvestment and Source Code Capital, and additional investments from existing shareholders including Gaorong Capital and Ant Group. 

Garnering so much attention shortly after their launch, it makes people wonder if these two rising stars can catch up with or even surpass their former employer. When ByteDance faces off against its former top talents, will the two departing protégés manage to outperform their old master?

Self-developed large models cannot compete against Seedance

While both are AI video generation products, Paiwo AI and LibTV follow completely different paths. LibTV functions more as an aggregation platform, still utilizing ByteDance's large model Seedance as well as Kuaishou's Kling, to generate designed products through well-structured workflow design and various skill combinations. Paiwo AI, on the other hand, uses its self-developed model Pixverse. Although it has recently integrated Seedance 2.0 as well, its core focus is clearly on model R&D.

Ultimately, at the fundamental technology level, the two products still represent a competition between Seedance and emerging large models.

Tech Planet conducted actual tests on both products: using the native Seedance 2.0 model on LibTV, and the self-developed Pixverse V6 model by PixVerse Tech on Paiwo AI. The test evaluated their understanding of prompts and image quality by having both generate the same video.

First, we gave the same prompt to both products, asking them to generate an anime-style video of Elon Musk watching a match at the FIFA World Cup hosted by the US, Canada, and Mexico.

The first version produced by Paiwo AI featured an overly slim character with an overly prominent jawline, which deviated significantly from Elon Musk's appearance. Moreover, the entire background did not look like a World Cup venue at all, but more like an esports arena.

We then sent a further prompt to the video, asking to change the scene to a more recognizable World Cup stadium and include a goal-scoring moment. This time the scene was correct, but the character's position was misplaced: the character's position suddenly shifted when the goal was scored, moving directly onto the football pitch. The character's reaction was also delayed, mismatched with the plot development.

We then combined the two prompts to regenerate the video. The scene and character movement were mostly accurate, but the character's facial features no longer matched the prompt at all. The character's clothes were changed to a jacket, which was neither the formal suit that fits Musk's daily style nor the jersey that fits the match scene. The problem of misplaced character positions still existed, showing very unstable performance.

In comparison, the video generated by LibTV presented a far more accurate character that closely matched Elon Musk's appearance. The character was wearing a jersey that fit the scene, and the character's facial expressions aligned with the plot, though the expressions were slightly exaggerated. The flag and gloves in the character's hand looked a bit out of place, and the camera angles were relatively monotonous.

After sending a further prompt to add a goal-scoring scene, the plot still maintained logical consistency.

We then asked both products to generate a video of a robot dancing ballet on stage. Paiwo AI still failed to accurately depict the scene, as it did not clearly show the stage element, while LibTV fully presented the stage setting.

From the product experience, for users with zero technical background, Pixverse V6 offers rich visuals and detailed designs, but it tends to overlook elements in the prompts and may generate illogical content. Users need to constantly adjust the prompts to improve the output precision. Seedance 2.0 executes instructions more completely with relatively smooth logic, but its performance on details remains unstable.

In terms of model usage, Paiwo AI only allows users to select one single model, but its operation method is simple with no technical barriers, targeting a broader user base. LibTV supports multi-model integration, with a very smooth operation workflow that better caters to professional users. The whole process from selecting skills or entering prompts to displaying the workflow and modifying nodes is very clear, which demonstrates its excellent implementation of model application.

Comparing the built-in templates and skills of the two products, Paiwo AI's templates are biased towards vertical-screen creative content, which fits the operational logic of short videos; LibTV's built-in skills are more focused on long-video styles, with a higher upper limit of professionalism.

Surpass the giant ByteDance?

Although both founders started their businesses after leaving ByteDance, their past experiences determined the different orientations of their products. Wang Changhu's technical background makes his product focus on model R&D, while Chen Mian's experience with Jianying has laid the foundation for his expertise in growth and commercialization.

At present, besides its self-developed large model, PixVerse Tech's core products are only Paiwo AI and its overseas version Pixverse, with over 150 million global users covering 177 countries and regions. However, its monthly active users have long stayed at around 15 million, with an annualized revenue exceeding $40 million.

Eanyo Tech, by contrast, owns multiple products including LiblibAI, StarFlow / Lovart, and LibTV, all of which have impressive performance data. LiblibAI has accumulated over 30 million users with annualized revenue exceeding $300 million. Lovart achieved an annualized revenue of $80 million only 5 months after its launch. LibTV gained over 100,000 users on its first launch day, and its monthly revenue increased by more than 13 times two months after launch.

The main advantage of Paiwo AI lies in its high level of autonomy at the underlying technology level, with low dependence on upstream resources and stronger risk resistance. Although its model is not yet perfect, the company has already mastered the core technology in its own hands.

However, the challenge also comes from the model itself. To compete with ByteDance's Dreamina, it must overcome key technical hurdles. Currently, there is still a clear gap between its self-developed model and Seedance, and it will take a considerable amount of time to reach the level of its former employer.

LibTV's core advantage is its pricing. In April 2026, due to scarce computing power, Dreamina adjusted its prices three consecutive times, leading to a nearly 6-fold increase in video generation costs. As a result, LibTV, which can also access Seedance 2.0, became an affordable alternative for many AI video practitioners. The premium version of LibTV costs 11,999 RMB per year, providing 66,000 credits monthly, while Dreamina's super membership costs 30,576 RMB per year with only 54,600 monthly credits, showing a significant price gap. However, LibTV's ability to maintain low prices relies heavily on substantial subsidies, which is not a long-term sustainable strategy.

As a model aggregation platform, LibTV fully depends on upstream large models at the fundamental technology level. Any changes to the upstream large models will directly impact its capital chain and product quality. Compared with products like Dreamina that are directly backed by ByteDance's self-developed large models, LibTV has weaker risk resistance and narrower development space.

For Chen Mian and Wang Changhu, who both left ByteDance to start their own businesses, ByteDance remains a formidable giant standing in their path. However, the future competition outcome is still uncertain. The AI video generation market is still expanding, and AI products have not yet entered the stock game stage, as the market is still in the bonus period of rapid expansion. There is ample room for new players, and anything is possible.

This article is from the WeChat official account "Xinyan Finance" (ID: tech621), written by Zhang Nongyi, and authorized for release by 36Kr.