Transfer over Sora 2, there’s a sizzling new AI video mannequin on the town.
In early February, ByteDance unveiled Seedance 2.0, a serious improve to its flagship video mannequin, which had beforehand remained pretty obscure. Its highly effective capabilities instantly shocked the AI ecosystem in China, even amongst audiences who had as soon as been skeptical of AI-generated video and seen the expertise primarily as a approach to produce slop.
Feng Ji, the founding father of Sport Science, the studio that developed China’s international hit online game Black Fantasy: Wukong, wrote on-line that he was “deeply shocked” by the mannequin’s skills and believed Seedance 2.0 would pose important challenges to China’s present copyright rules and content material moderation methods. Pan Tianhong, who leads a Chinese language skilled video manufacturing studio with over 15 million followers on social media, posted a video during which he mentioned Seedance 2.0 is considerably higher than any video-making fashions that got here earlier than it. “It thinks like a director,” Pan mentioned.
Nonetheless, most individuals can’t get their fingers on the mannequin at this second as a result of entry stays pretty restricted. As of this week, ByteDance is just permitting current customers of its consumer-facing AI apps in China—the preferred one is the chatbot app Doubao, however the firm additionally has a complicated constellation of lesser-known apps like Jimeng, Xiaoyunque, and Spark—to expertise Seedance 2.0. All these apps are for the Chinese language home market solely, stopping individuals outdoors the nation from testing the mannequin themselves. (The restrictions have prompted some savvy individuals in China to resell their ByteDance accounts to keen early AI adopters abroad.)
However there are indicators that the mannequin may turn into extra accessible quickly. This week, ByteDance up to date its API platform and disclosed the proposed pricing of Seedance 2.0: A 15-second video, the longest it could possibly generate proper now, would value barely greater than $2 to make, Chinese language publication IT Dwelling estimated. ByteDance nonetheless hasn’t opened up API entry to third-party builders, however that ought to be on the horizon.
Afra Wang, creator of the Substack publication Concurrent and a detailed observer of the US-China AI panorama, tells me that Seedance 2.0 is one other fascinating instance of how the 2 nations have taken diverging paths. Even earlier than the discharge of Seedance 2.0, a number of the most established video-making AI instruments on the earth, similar to Kling AI, have been developed by Chinese language corporations. “China hasn’t produced any respectable AI coding software, which is why Chinese language persons are all depending on Claude Code or Codex; however in the case of video AI, China is miles forward of the US,” Wang says.
However all of the hype apart, Seedance is operating into two severe issues. Weeks into its launch, ByteDance is going through a compute bottleneck that’s inflicting the mannequin to take hours to generate a single video. In the meantime, main film studios, together with Disney, Netflix, and Paramount, have all despatched ByteDance cease-and-desist letters alleging that Seedance 2.0’s outputs are infringing on their copyrighted works. ByteDance didn’t instantly return a request for remark.
The Bandwidth Downside
Even in case you get entry to a ByteDance AI app, it’s nonetheless removed from straightforward to generate a video with Seedance 2.0, as a result of too many individuals are attempting to do the identical factor, and ByteDance has but to offer sufficient compute sources for everybody.
After I tried to create a clip with one in all ByteDance’s apps this week, it advised me that I used to be quantity 90,985 within the queue, and it might take about 4 hours to generate a five-second video. After ready for 2 hours, the app advised me I now had six extra to go. At that time, I made a decision to only go to mattress.
