While few dispute that China is narrowing the gap with the United States in developing artificial intelligence (AI) models, there is one area where Chinese firms are already the undisputed leaders: video.

China has seen a proliferation of advanced generative AI video models – systems trained on visual data to generate or edit video content based on user inputs such as text prompts, images and audio.

The country’s vast short-video ecosystems and looser copyright rules have helped give Chinese companies the advantage on AI video models. This is one arena where computing capacity, on which the US has a lead, does not dictate success.

Chinese tech giants ByteDance and Alibaba and start-up MiniMax dominate global AI video leader boards, with a model from Google their only consistent competitor.

Expertise in developing AI video models, especially multimodal models that process multiple types of data input, has helped China cement its lead in physical AI such as robots and autonomous vehicles. To be effective, a video model needs some understanding of physics – a sense of motion, causality and how objects behave.

These are also essential for building the AI world models – machine learning systems that build an internal representation of an environment – which companies use to train and operate robots and self-driving cars.

World model development is a sphere in which China has a competitive edge (see China Future Tech briefing), having been quick to integrate AI in physical settings including factories. This has allowed the accumulation of masses of real-world data to improve models’ performance.

Chinese firms’ prowess in building AI video models could help the country preserve its edge in world models, which leading AI video firms like Beijing-based Shengshu Tech have also begun releasing.

AI video generation tools allow someone with no camera or editing skills to create a professional-looking video using a text prompt, images, a script or existing audio and video clips – or combinations of them – and to do so quickly and cheaply.

For example, a toy salesman at Yiwu International Trading Market in Zhejiang province can post a sales video in an online marketplace featuring his avatar to introduce products in various foreign languages. Yiwu is the world’s largest wholesale market for small commodities.

A property agent with an empty house to sell but no budget to rent furniture or commission three-dimensional renderings of furnished rooms can create a video with 3D digital furniture to draw more buyer interest.

Instead of doing a mass blast of a sales video, models help to create a templated video framework with custom elements such as a prospective customer’s name and relevant use case for personalised outreach.

The models can also be used to quickly turn a long-form marketing video into a week’s worth of platform-native clips – square for LinkedIn and X, horizontal for YouTube, vertical for TikTok and Reels.

The result is radically reduced time and costs. AI video production and editing platform Invideo says the average cost per finished minute of AI-produced video is between US$315 and US$750, based on five videos of varying lengths shot using leading AI models. That compares to between US$1,000 and US$50,000 per finished minute for traditional video production, it says. Invideo says AI can produce a two-minute commercial in as little as three days compared to two months if produced by conventional methods.

Around the world, personal and business users are turning to Chinese AI video models. They regularly occupy five or more of the top 10 places in Artificial Analysis’ benchmark rankings of the most powerful AI text-to-video, image-to-video and video editing tools. The industry evaluation platform based its rankings on blind testing by a panel of evaluators.

In July, Neill Blomkamp, the South African director behind science-fiction feature films , and , posted a 13-minute sci-fi short film called on YouTube. It was generated entirely using the Seedance 2.0 AI video generation model from Beijing-based ByteDance, owner of TikTok.

Seedance 2.0, released in February, quickly drew praise for its ability to generate lifelike characters and for the precise editing control it offers.

Blomkamp directed shot by shot using text prompts. In a post on X he said he now wants to “tackle a full feature in this format”.

He won’t be the first to do so. , a 95-minute US production created using Seedance 2.0 and billed as the world’s first AI-generated feature film, premiered at the Cannes Film Festival this year.

On Friday, the first AI-produced feature to get a cinematic run opened in China. Made using more than 1.2 million source images, runs for 100 minutes and combines China’s ancient Shu civilisation with a futuristic sci-fi setting, tech news website Technode reported.

Soon after ’s release, ByteDance launched Seedance 2.5, which offers improved textures, lighting and skin detail and allows users to create scenes with multiple characters and camera angles, according to tech website The Decoder.

In August, Hangzhou-based Alibaba released its latest video generation model, Wan 3.0. It also accepts various multimedia inputs including documents in PDF and other formats, web pages and PowerPoint slides, while matching Seedance 2.5 with a maximum single-clip length of 30 seconds. Alibaba owns the South China Morning Post.

At the time of writing, Wan 3.0 occupied top spot in Artificial Analysis’ ranking of AI text-to-video models, with Seedance 2.5 in fifth place and Seedance 2.0 in ninth. Two versions of the MiniMax H3 model occupied fourth and seventh spots. Utopai X, from California-based Utopai Studios but post-trained on MiniMax H3, was in second place.

Released in July, MiniMax H3, from the eponymous Shanghai-based AI company, is open-weight for most users, unlike the Alibaba and ByteDance models. Weights are the underlying parameters that encode an AI’s intelligence. Open-weight models can be downloaded and fine-tuned.

The same Chinese models also make Artificial Analysis’ top 10 for video editing, along with a version of the multimodal Kling 3.0 model from Beijing-based Kuaishou. Kling models are noted for ease of use and complex editing functions such as swapping in characters, editing out backgrounds and retouching parts of a video while keeping everything else consistent. Also ranked for video editing is a version of HappyHorse from Alibaba, designed to produce cinematic short clips.

Most of these models also rank among the top 10 image-to-video models. The only model wholly developed in the West that rivals them in all three categories is Google’s Gemini Omni Flash 1.1.

One factor behind the success of Chinese AI video models is the microdramas that have exploded in popularity in China since 2018. Aimed at mobile phone users, with dozens of episodes one to two minutes long, they feature fast-paced storylines and frequent plot twists.

They screen on social media apps, including Kuaishou and ByteDance’s Douyin – the Chinese TikTok. That gave both companies a head start on AI video models, since the proliferation of these drama series provided unique data sets for video model training.

Another factor is the volume of creators attracted to Chinese AI video models. The content these users generated fed back into the models’ iteration.

Chinese AI video models improved so rapidly that the companies making microdramas were soon racing to use them to reduce their costs and production times.

In January this year, a new AI-generated microdrama went live in China every 90 seconds, tech news website The Next Web reported. In March, 50,000 AI-native titles were added to Douyin alone, it said, with their production costs just a 10th those of a live-action shoot. In the first half of the year, Douyin recorded 500 billion views for 220,000 new AI-generated microdramas, CNN reported, citing market research firm DataEye.

The training of generative AI models, including video models, on vast amounts of copyrighted material, and the content created using them, has prompted lawsuits around the world, including in China.

A Chinese regulatory approach of “moderate leniency” has emerged that emphasises control over AI-generated content while adopting a more permissive stance towards model training, Chinese legal scholars Li You, of East China University of Political Science and Law in Shanghai, and Han Luo, of the University of International Business and Law in Beijing, argued in a legal journal paper published in 2025.

Comparatively flexible copyright enforcement in China gave domestic firms more runway to test and commercialise their models, said Wang Zebin, an investment manager at private equity firm JG Investment.

US AI developers faced “much more potential risk and liability” over intellectual property than counterparts in China, Kyle Chan, a fellow at the Brookings Institution’s John L Thornton China Centre, told the SCMP.

Courts in most jurisdictions have shown a reluctance to recognise AI-generated content as being of human authorship and subject to copyright protection, Tsinghua University legal scholar Shujie Feng explained in another 2025 journal article, but Chinese courts have, in numerous judgments since 2018, reached the opposite conclusion.

In one widely cited judgment, the Beijing Internet Court ruled in 2023 that an AI-generated image enjoyed copyright protection in view of the prompts and parameters the user adopted to produce it.

Chinese judges have granted copyright for AI-generated work where they have been satisfied that it reflects human originality and intellectual effort, Paolo Beconcini, head of China intellectual property for US-based global law firm Squire Patton Boggs, explained in a March 2025 LinkedIn post. Rulings in copyright cases against the developers of AI models have been less clear-cut.

Judges in two 2024 actions involving platforms that allowed users to generate images similar to the fictional alien superheroes drew opposite conclusions. A court in the southern city of Guangzhou found that training an AI model on copyrighted material without authorisation and allowing users to output substantially similar material infringed on copyright. A court in eastern Hangzhou found another platform had not directly infringed on copyright.

Last year a court in Shanghai cleared an AI large model developer of copyright infringement over an image of a character from an animated series, but ruled a user of the model infringed on copyright.