iFlytek unveils real-time hyper-realistic digital human tech

Chinese artificial intelligence pioneer iFlytek on Thursday unveiled its "Ultra-Realistic Digital Human Real-Time Generation Technology" on Thursday. The company describes the technology as reducing digital human creation work from hours of video capture to a single photograph, enabling more natural and complex interactions.

iFlytek's latest move underscores the accelerating shift from passive visual avatars to active, intelligent digital workforces, as generative AI reshapes the economics of virtual interaction.

Traditionally, producing a lifelike digital human required about two hours of human video recording and lengthy training cycles. The new system, demonstrated at the launch event, generates a digital human instantly from one still image, complete with natural facial expressions, gestures, and lip-sync.

More notably, the real-time avatar can perform intricate actions such as picking up objects, drinking, singing, and dancing, delivering smoother and more immersive human-machine interaction, the company said.

According to iFlytek, the achievement rests on several core technologies: iFlytek's Spark voice large language model powers fluent dialogue in general knowledge domains; curated vertical-industry knowledge bases reduce hallucination rates and provide precise professional answers; and customizable personas with long-term memory enable the digital human to maintain consistent character traits and accumulate interaction history over time.

Current deployment scenarios span brand live-streaming, customer service, and unmanned retail stores. As large-model adoption lowers cost barriers, digital humans are transitioning from a premium tool for large enterprises to a standard asset for a broader range of businesses, iFlytek said.