China’s generative AI scene is taking shape across consumer apps, research labs and commercial products. The tools reflect local language and culture, but their creators also face questions about bias, regulation, business models and access to computing power.
Models built for local audiences
Baidu’s ERNIE-ViLG is a text-to-image model trained on a data set of 145 million Chinese image-text pairs. A test involving shumai illustrated how cultural context can affect what an image generator produces: Stable Diffusion captured a Chinatown restaurant atmosphere, while ERNIE-ViLG made a type of shumai more commonly seen in eastern China than the Cantonese version.
The comparison points to a broader challenge. Models learn from the images and captions available to them, and those collections can represent some communities or traditions better than others. A system may produce a convincing scene while still missing the details that matter to people familiar with its subject.
Tencent’s Different Dimension Me, which turns photos into anime characters, encountered another kind of gap. It gained unexpected users in anime-loving regions such as South America, but users found that it failed to identify black and plus-size individuals. The source connects this problem to the noticeable absence of those groups in Japanese anime.
Research is also happening beyond commercial tech firms. Taiyi, developed by the IDEA research lab, is an open source model trained on 20 million filtered Chinese image-text pairs and has one billion parameters. The lab, led by Harry Shum, is backed by local governments and based in Shenzhen.
Rules shape how people use the tools
China’s approach to generative AI is shaped by regulation as well as training data. The source reports that Baidu’s image model filters politically sensitive keywords. It also describes measures for “deep synthesis tech,” a term covering systems that generate text, images, audio, video and virtual scenes.
Users of generative AI apps are asked to verify their names, making prompts traceable to real identities. That can discourage some behavior. The rules also explicitly prohibit generating and spreading AI-created fake news, though the source notes that service providers would be responsible for implementation.
These requirements raise a difficult policy question: how can rules reduce harmful uses while leaving room for experimentation? Yoav Shoham, co-founder of AI21 Labs, described regulation as an area where people are still learning together. He argued that the work should involve technologists, the public sector, social scientists, affected people and government.
Creators and companies look for revenue
Some people in China are using image generators to make emojis and wallpapers for social media, where they can earn advertising revenue or charge for downloads. Others sell prompts or teach people how to use them. These examples show how a tool can become a source of income even for users without technical backgrounds.
Businesses are exploring more targeted uses. Light fiction writers can make illustrations for their work, while manufacturers can generate batches of designs for products such as T-shirts, press-on nails and prints. The source says this could lower design costs and shorten production cycles.
During the pandemic, e-commerce sellers struggled to find foreign models while China’s borders were shut. Sequoia China–backed Surreal, later renamed to Movio, and Hillhouse-backed ZMO.ai worked on algorithms to generate fashion models of different shapes, colors and races.
Still, commercial potential does not guarantee easy growth. Founders cited weaker willingness among Chinese enterprise customers to pay for SaaS than customers in developed economies. One anonymous founder also described intense competition, where rivals could copy source code and use customer support staff to compete for users. Shi Yi, founder and CEO of sales intelligence startup FlashCloud, said Chinese tech firms often focus on applications and quick returns.
Chip access remains a concern
In September, the U.S. government imposed export controls on high-end AI chips. The source says companies focused on basic research may face longer and more expensive computing when using less powerful chips, even though many startups building applications do not need the same level of hardware.
Baidu executive vice president and head of AI Cloud Group Dou Shen said the controls’ impact on the company’s AI business was “limited.” He cited existing chip supplies for near-term needs and Baidu’s own Kunlun AI chip for the medium to longer term. The source reports Baidu’s claim that using Kunlun chips in large language models improved efficiency for text and image recognition tasks by 40% and reduced total cost by 20% to 30%.
Whether Kunlun and other domestic chips can help China gain an advantage remains uncertain. For now, the country’s generative AI development is being shaped by a mix of local products, cultural and data challenges, regulation, commercial pressures and hardware constraints.