ByteDance has released Seedance 2.0, a powerful new AI video model that has impressed industry professionals in China with its advanced, director-like capabilities. However, access is currently restricted to users of ByteDance's domestic Chinese apps, though API pricing suggests a broader release may be coming.
The model highlights China's competitive edge in AI-generated video compared to the US, according to observers. Despite the hype, Seedance 2.0 faces significant challenges including severe compute bottlenecks causing long wait times and cease-and-desist letters from major Hollywood studios over alleged copyright infringement.
Main Topics: Seedance 2.0's release and capabilities; its restricted access and potential future availability; China's position in the AI video landscape; the model's current problems with computational bandwidth and copyright disputes.
Move over Sora 2, there’s a hot new AI video model in town.
In early February, ByteDance unveiled Seedance 2.0, a major upgrade to its flagship video model, which had previously remained fairly obscure. Its powerful capabilities immediately shocked the AI ecosystem in China, even among audiences who had once been skeptical of AI-generated video and viewed the technology mainly as a way to produce slop.
Feng Ji, the founder of Game Science, the studio that developed China’s global hit video game Black Myth: Wukong, wrote online that he was “deeply shocked” by the model’s abilities and believed Seedance 2.0 would pose significant challenges to China’s current copyright regulations and content moderation systems. Pan Tianhong, who leads a Chinese professional video production studio with over 15 million followers on social media, posted a video in which he said Seedance 2.0 is significantly better than any video-making models that came before it. “It thinks like a director,” Pan said.
However, most people can’t get their hands on the model at this moment because access remains fairly restricted. As of this week, ByteDance is only allowing existing users of its consumer-facing AI apps in China—the most popular one is the chatbot app Doubao, but the company also has a confusing constellation of lesser-known apps like Jimeng, Xiaoyunque, and Spark—to experience Seedance 2.0. All these apps are for the Chinese domestic market only, preventing people outside the country from testing the model themselves. (The restrictions have prompted some savvy people in China to resell their ByteDance accounts to eager early AI adopters overseas.)
But there are signs that the model might become more accessible soon. This week, ByteDance updated its API platform and disclosed the proposed pricing of Seedance 2.0: A 15-second video, the longest it can generate right now, would cost slightly more than $2 to make, Chinese publication IT Home estimated. ByteDance still hasn’t opened up API access to third-party developers, but that should be on the horizon.
Afra Wang, author of the Substack newsletter Concurrent and a close observer of the US-China AI landscape, tells me that Seedance 2.0 is another interesting example of how the two countries have taken diverging paths. Even before the release of Seedance 2.0, some of the most established video-making AI tools in the world, such as Kling AI, were developed by Chinese companies. “China hasn’t produced any decent AI coding tool, which is why Chinese people are all dependent on Claude Code or Codex; but when it comes to video AI, China is miles ahead of the US,” Wang says.
But all the hype aside, Seedance is running into two serious problems. Weeks into its release, ByteDance is facing a compute bottleneck that is causing the model to take hours to generate a single video. Meanwhile, major movie studios, including Disney, Netflix, and Paramount, have all sent ByteDance cease-and-desist letters alleging that Seedance 2.0’s outputs are infringing on their copyrighted works. ByteDance did not immediately return a request for comment.
The Bandwidth Problem
Even if you get access to a ByteDance AI app, it’s still far from easy to generate a video with Seedance 2.0, because too many people are trying to do the same thing, and ByteDance has yet to provide enough compute resources for everyone.
When I tried to create a clip with one of ByteDance’s apps this week, it told me that I was number 90,985 in the queue, and it would take about four hours to generate a five-second video. After waiting for two hours, the app told me I now had six more to go. At that point, I decided to just go to bed.
Part of why the wait time got longer could be because ByteDance gives priority to people who pay for its monthly AI subscription programs. (The highest tier costs over $70 a month, which is prohibitively expensive in China, where people are used to most apps being free.) “Even if you are a paid member, you’d still need to wait in the line for a long time,” Wang notes. Despite having a ByteDance account, Wang says she hasn’t generated any videos with Seedance 2.0 yet because she couldn’t get over the frictions of adoption.
Instead, Wang went on social media for tips, where other Seedance users have shared strategies for reducing the waiting time as much as possible, such as asking the model to generate shorter videos, sending requests after midnight when fewer people are online, and even renting premium accounts from other people.
For now, the biggest bottleneck for ByteDance seems to be finding and allocating enough GPUs for its video-making product. In general, generating high-quality videos uses far more compute resources than generating text responses. Even though ByteDance is one of the Chinese tech companies with the deepest pockets, it still doesn’t have access to the number of data centers that its Western AI competitors do, and it specifically doesn’t seem ready for wide adoption of Seedance 2.0 at the moment.
People on social media have also been sharing another common frustration: After waiting hours for a video to reach 99 percent completion, the model enters a final review stage where the app sometimes decides the result fails content checks and refuses to release it, sending users back to the end of the queue.
Muted Reactions
In addition to widespread enthusiasm from the general public, it’s been interesting to see the entertainment industry in China react much more positively to Seedance 2.0 than Hollywood has been reacting to generative AI broadly.
On February 16, Jia Zhangke, a Cannes-winning Chinese film director, posted a video that he said he “collaborated” on with ByteDance’s Doubao chatbot. The five-minute clip generated by Seedance 2.0 features two AI-generated avatars of Jia talking to each other and remaking classic scenes from the director’s movies. “I don’t worry about whether technology will replace movies,” Jia wrote on the Chinese social media platform Weibo. “What actually matters is how the technology is used by people.”
Jia’s view contrasts sharply with the dominant reaction to AI in Hollywood. Celebrity directors and actors have remained mostly silent on the topic, and large production houses have shied away from using AI, at least publicly. In China, though, the entertainment industry has been more willing to embrace the technology. Seedance 2.0 was used to generate backdrops during this year’s Spring Festival Gala, the highly anticipated annual broadcast where corporate participation is often seen as one of the strongest stamps of approval from the Chinese government.
“When a famous director in China jumps onto the AI bandwagon, they aren’t getting much condemnation from their peers. But it’d look totally different in the United States,” says Wang.
One issue, however, is that intellectual property protections are much less developed in China than in the US, and it’s hard for Chinese moviemakers to sue AI companies or block them from using their content. That legal reality has, in turn, led consumers and creatives to normalize alleged infringement of intellectual property. When Pan Tianhong, the Chinese video creator, tested Seedance 2.0, he was startled to realize the model could mimic his precise speaking voice, which he neither explicitly agreed to nor was compensated for. Yet, he was willing to brush it off as the likely result of some obscure terms and conditions that he agreed to when he signed up for social media.
Meanwhile, the looser IP protections also mean Chinese creators are able to use AI to generate more familiar characters, potentially making existing fandoms excited to share AI-generated content. Users are already using video AI to generate scenes from famous TV series and movies in China. That kind of content is fueling the popularity of Chinese AI models.
But the moment that kind of content begins scaling globally, it could become a significant legal liability. Even during the short period between when ByteDance released Seedance 2.0 and when major studios began accusing the company of generating copyrighted materials, I saw videos on X in which Wolverine fights Hulk and Tom Cruise battles Brad Pitt. But things went south very, very fast, and I even saw a dance-off between Michael Jackson and Hitler. (Yes, that video actually exists.)
Seedance 2.0 is a glimpse of where generative AI in China is heading: It’s getting increasingly capable and being adopted by real professionals, but it’s still hampered by infrastructure constraints and legal risks.
This is an edition of Zeyi Yang and Louise Matsakis’ Made in China newsletter. Read previous newsletters here.
Meridiano
