Chinese AI groups pull ahead of US rivals in video generation race

{"text":[[{"start":7.95,"text":"Chinese artificial intelligence groups have moved ahead of US rivals in video generation, a key battleground in generative AI in which there is rapid uptake across advertising, ecommerce and entertainment."}],[{"start":21.2,"text":"Companies such as Beijing-based ByteDance and Kuaishou are training systems on vast libraries of short-form video, giving them an edge over American competitors."}],[{"start":31.9,"text":"The shift marks a divergence in the AI race: while OpenAI, Google and Anthropic still dominate large language models and areas such as coding, their video tools lag behind Chinese offerings in quality and usability, according to developers and multiple usage leader boards."}],[{"start":null,"text":"

"}],[{"start":null,"text":""}],[{"start":49.9,"text":"Training video models requires vast amounts of high-quality footage, an area where Chinese platforms have an edge through their ownership of short-video apps, such as TikTok, and the data they generate. Some experts argued Chinese groups had also been more aggressive on accessing copyrighted material. "}],[{"start":68.6,"text":"“Most of the American models that we’ve tried are not very good at video generation,” said Ben Chiang, founder of Director AI, a start-up that produces AI-generated content such as cartoons and short dramas. He said stricter content controls constrained US tools and often produced less realistic outputs."}],[{"start":88.44999999999999,"text":"Director AI mainly uses Kling, developed by Kuaishou, while also switching between ByteDance’s Seedance 2.0 and Hailuo from start-up MiniMax, depending on the task and cost. "}],[{"start":100.44999999999999,"text":"“It comes down to quality and how well the model follows the prompt,” Chiang said, adding that recent advances had made the tools smoother to use, particularly in synchronising audio and stabilising voices."}],[{"start":112.89999999999999,"text":"For creators, the improvements are already reshaping production. “Seedance 2.0 has been a game-changer for my creative freedom,” said George Won, an independent AI filmmaker and editor in Tbilisi, Georgia."}],[{"start":126.64999999999999,"text":"“It handles aggressive camera angles and speed without losing the character’s face or lighting contrast. Most AI models start to wobble or drift when things move fast,” he added."}],[{"start":null,"text":""}],[{"start":138.1,"text":"Kling, Seedance 2.0, and HappyHorse 1.0 all scored highly in Arena's ranking of the best video models, compiled from votes by users on the independent platform. Such assessments are viewed as more reliable than most model benchmarking systems, which AI labs optimise for and are not necessarily accurate reflections of their real-world performance. "}],[{"start":160.1,"text":"Google’s Veo 3 model is also competitive, which experts say has an edge from its access to YouTube video footage, but has more safeguards and limitations on content for developers. "}],[{"start":171.75,"text":"The progress is underpinned by access to vast volumes of proprietary video data. ByteDance and Kuaishou operate some of the world’s largest short-video platforms, giving them a training advantage that is difficult for competitors to replicate — particularly as video content, unlike text, cannot easily be scraped at scale. "}],[{"start":192.95,"text":"This week, Kuaishou announced that it was exploring options to spin off the Kling business, and potentially pursue a separate listing to capitalise on its fast-growing AI video business."}],[{"start":203.39999999999998,"text":"The use of such material has also drawn scrutiny. ByteDance has faced legal threats over alleged copyright violations by enabling users to create videos with characters, including from Marvel movies and the South Park television series, without proper permission. The Chinese internet group has pledged to increase protections."}],[{"start":222.99999999999997,"text":"Several developers and creators said looser content restrictions made Chinese models easier to use in practice. Chiang said US tools would “constantly run into errors”, denying requests for violating usage terms without further explanation."}],[{"start":237.54999999999998,"text":"For individual creators, Chinese platforms are often cheaper and more flexible than western alternatives, offering subscription or pay-as-you-go systems that enable rapid experimentation. High demand for Seedance 2.0 since February, however, has led to restricted usage and long wait times for some users."}],[{"start":257.09999999999997,"text":"For enterprise customers, ByteDance has adopted a different approach. For some US-based clients, it has required large upfront commitments, asking customers to pay about $2mn for access and credit allocation, according to people with knowledge of the matter. ByteDance did not respond to a request for comment."}],[{"start":276.9,"text":"Some users circumvent these barriers by accessing the models through third-party platforms such as BeHooked, which repackage them with alternative pricing structures."}],[{"start":null,"text":""}],[{"start":null,"text":""}],[{"start":286.54999999999995,"text":"Generating video requires vastly more tokens — the units of data processed by AI models — than text or audio, making large-scale deployment expensive. In March, OpenAI discontinued its Sora video-generation model in part because of the high computing costs."}],[{"start":303.65,"text":"For companies building on top of the technology, the recent leap in quality is opening commercial opportunities."}],[{"start":310.04999999999995,"text":"“Before, the videos were cringey and robotic — they couldn’t meet brand standards,” said Vincent Yang, chief executive of Firework, which provides video infrastructure for ecommerce websites. “Now we’re at the point where you can’t tell if it’s AI or human.”"}],[{"start":325.65,"text":"Yang said the technology was already reshaping advertising by enabling brands to generate videos at a scale previously impossible."}],[{"start":334.7,"text":"“One retailer asked us to create 100,000 videos for its product pages,” he said. “Without AI, that would have been prohibitively expensive. Now, every product can have its own video, and even multiple versions tailored to different customers.”"}],[{"start":350.05,"text":"Additional reporting by Cristina Criddle in San Francisco "}],[{"start":362.35,"text":""}]],"url":"https://audio.ftcn.net.cn/album/a_1778999080_5881.mp3"}