Unique Research · Insight / longform · May 6, 2023 · Full rewrite — not a stub · Interviews with 6 digital-human vendors
Digital-human tech is approaching mass adoption. Brands fled expensive, “collapse-risk” celebrity endorsements toward 100%-controllable digital humans. Twin forms: virtual humans (fully synthetic) vs digital twins/avatars of real people (e.g. Daoying’s AI twin of ophthalmologist Tao Yong). GenAI (text/image/audio → 3D, video, animation, music, poetry) gives digital humans a “brain” for lifelike interaction — personal twins within ~2–3 years feel plausible.
Unique Research interviewed six China vendors: Yizhi Intelligence, Daoying Yousheng, Yuanjing Tech, Tuoyuan Wisdom, Chuangxiang Dimension, Zhouyu Tech.
02. Tech stacks
Bottom layers cluster around NLP, multimodal interaction (speech/gesture/touch/gaze), foundation models, and XR (AR/VR/MR).
Yizhi (ZJU AI Institute roots) — CTO Li Yifu: started NLP → smart CS; pivoted ~2yrs ago to multimodal virtual humans (speech+text + photoreal + dynamic video). Claims lead in real-person drive + lip-sync; consumer-domain speech/semantic models + interaction modules; strong PMF help from existing consumer clients.
Yuanjing (Tianyu Digital strategic) — founder Wang Zhiwu: NLP semantic understanding, pretrained multimodal LMs, algorithm kits on Yuanxiang intelligent cloud — 1+X product system lowering application barriers; focus vertical scenes over generic tools.
Daoying — CEO Xiao Shuo: multimodal neural rendering engine (text/audio/image + motion drive). B2B solutions, UI, APIs for A/V production. Two schools: classic 3D + UE/Unity (games/film) vs Daoying’s no-traditional-modeling cloud pipeline — lower cost/time, no foreign monopoly render engines, standardized multimodal content.
Tuoyuan — co-founder / Yuanshenshen GM Huang Weipeng: multimodal LM + cognitive/causal/mental models for content gen + virtual–real interaction; products in real-person twins, AI co-creation, enterprise digital services (100+ clients). Live: parse danmaku → answer; short-video AI copy; 1:1 ASR→NLP→TTS + face/gesture. Claims AI livestream GMV >3× human; 7×24 “super studios” for SME merchants short on budget/talent.
Chuangxiang — CEO Wu Wei: design→model→drive→ops pipeline accelerated by AI components. Old stack = knowledge graph + speech/action libraries (closed Q set); with LLMs + TTS + speech-to-motion + expression/action libs, assemble Lego-like drivers that answer anything. Expects 5–10× per-role productivity in 3–5yrs; AIGC era starts with tertiary (services/entertainment/content) then factory floors — competitors at 3–5× output/person will cull laggards.
Zhouyu — founder Zhu Feng: interaction = vision/hearing/touch (touch ultra-early). Hearing: Xiaoice/iFlytek/Ali push TTS fluency + emotion; still weak on prosody/volume/emotion nuance. Vision: classic mesh+texture CG shaken by NeRF → neural rendering paths. Sees Plugin Store as next App Store once industry knowledge plugs mature.
03–04. Commercial rise & genAI “bringing them alive”
Pain today: tiny corpora → repetitive answers; stiff face/gesture; weak emotion in voice. Yuanjing invested expert teams + AIGC/pretrained LMs + action DBs toward ~90%+ realism (100% = industry goal). ChatGPT-class LMs diversify answers + persona tone; audio2face replaces stiff text→lip pipelines for emotion-driven faces.
Metaverse upside: performance/art unconstrained once genAI matures. SME livestream supply/demand mismatch is a near-term killer app for digital-human + AIGC stacks.
13 CN images; attach on ingest. Author: Tan Qiuping; review Ski.
— END —