---
title: "Baidu's YiJing Pushes Digital Human Video to US$0.03 a Second — and Adds Micro-Expressions to Kill the 'AI Face"
date: 2026-11-18
category: Generative Media
site: NeuroAI
canonical: https://neuroai.site/a/na-aigc-baidu-yijing-video-02-yuan
language: en
---

# Baidu's YiJing Pushes Digital Human Video to US$0.03 a Second — and Adds Micro-Expressions to Kill the 'AI Face

> At Baidu World 2026, the YiJing digital human platform unveiled a self-developed video generation model with three upgrades: generation cost cut to 0.2 yuan (about US$0.03) per second, AU-level micro-expression control at 8–9 expressions per second, and full-frame real-time interaction under two seconds of latency.

At Baidu World 2026 on November 12, one of the most-watched videos was an interview where two versions of Robin Li (李彦宏) appeared on screen at once — the actual founder and a digital human replica conducting the conversation. The segment was itself AI-produced by YiJing (百度一镜), Baidu's digital human platform, and it served as the demo reel for the platform's biggest announcement of the day: a self-developed video generation model with three capability upgrades aimed squarely at making digital humans cheap enough, and lifelike enough, for mass commercial use.

## The three upgrades, unpacked

Baidu distinguished architect Yang Yi (杨羿) broke the release into three technical layers:

**1. Video generation at 0.2 yuan per second — about US$0.03.** Using hybrid diffusion distillation and sparsification modeling to compress inference, the model drives generation cost down to 0.2 yuan (about US$0.03) per second of video. A "Fast" version achieves one-to-one speed: a 10-second video produces in 10 seconds. For context on what that price point unlocks, an 80-minute standardized AI-produced drama now computes to a five-figure-yuan production budget (roughly US$1,400 to US$14,000) — and Baidu disclosed that its digital human livestream setup modeled on Luo Yonghao (罗永浩) has fallen from nearly 10,000 yuan per month at launch to a four-figure monthly cost, roughly US$140.

**2. Micro-expressions as a separate large model.** The Wenxin YiJing micro-expression large model delivers AU-level facial control — Action Units, the fine-grained coding system used in facial expression research — rendering up to 8–9 distinct micro-expressions per second, coordinated by a quality agent scoring over 200 sub-dimensions. The target is the stiffness problem: Baidu Group vice president Ping Xiaoli (平晓黎) told reporters that much of a digital human's "AI feel" comes from speech pauses, breathing gaps, and rhythm artifacts, so the fix is combining facial control with more colloquial, rhythmic speech.

**3. Full-frame real-time interaction.** Earlier interactive digital humans generated only a local region of the frame. The upgraded model streams the full frame in real time with a response latency under 2 seconds, using an autoregressive architecture plus retrieval-based memory to sustain interaction of unlimited length — the difference between a looping avatar and a service agent that can take customer questions indefinitely, as in the 4S car-dealership scenario Baidu showed, where the digital human recalls prior conversations and recommends test drives across languages.

## The commercial wrapper: short comics-drama Studio and four subsidy programs

On top of the three capabilities, Baidu opened the Short Comics-Drama Studio (短漫剧Studio), built on a two-agent pipeline: a script agent that ingests million-word web novels, extracts characters and plot arcs, splits episodes, and places reversal beats; and a production agent that carries storyboards, visuals, and voice through to finished episodes. A one-click style-transfer (一键转绘) function adapts plots, characters, and scenes for overseas markets — reframing short-drama localization from a heavy, per-frame workflow into a batch operation.

The platform also announced four ecosystem programs: 300 yuan (about US$42) in video credits per new user, 100,000 free digital human livestream slots for small merchants, 100 million yuan (about US$14 million) per year in content-creator incentives, and rebates of up to 60% for service providers.

## What the vendor numbers say — and don't

Baidu reports YiJing now covers more than 30 industries, works with 200,000 partnered livestream hosts, and has generated over 1 billion pieces of digital human content, while ranking first in Frost & Sullivan's digital human market share and overall strength assessments. Case figures include an e-commerce solution that cut costs 83% and lifted conversion 29%, and two international sportswear brands' overseas livestream rooms posting GMV growth above 40%. A Coca-Cola TVC case was presented as replacing a production budget of hundreds of thousands to millions of yuan (roughly US$14,000 to US$140,000).

These are Baidu's own or its clients' figures, presented at Baidu's own conference. The direction they point — digital humans moving from bespoke, studio-priced assets to commodity infrastructure — is plausible and consistent with the cost curve; the specific percentages are unaudited.

## Honest limitations

The 0.2-yuan-per-second (about US$0.03) figure is a headline rate, and real-world spending depends on generation retries, editing, and what Baidu counts as billable seconds — pricing details beyond the launch event remain undisclosed. Micro-expression control improves perceived realism, but "8–9 micro-expressions per second" is a vendor metric with no independent protocol, and perceived uncanny-valley effects vary by audience. The sub-2-second latency figure presumably applies under specific network and compute conditions, not every deployment. Frost & Sullivan rankings are frequently commissioned reports. And the Fast version's 1:1 speed applies to a defined video length; longer or higher-fidelity output will not run at that ratio. None of the released materials addresses content-rights questions around AI-produced presenters, an increasingly live legal area in China and abroad.

## What readers can do now

- **Price your own workload before believing the headline.** If you run video content operations, compute your current per-second cost and compare against US$0.03/s plus platform fees; the delta, not the rate, decides whether to migrate.

- **Test the uncanny-valley fix on your audience.** Run A/B exposure of the micro-expression model against a standard avatar with real viewers in your market — expression realism preferences differ sharply across demographics.

- **Pilot the Studio pipeline on one novel.** The script-agent-to-finished-episode flow is most credible on material you already own the rights to; use it to measure actual episode throughput and editing overhead.

- **Watch who else crosses the US$0.03 line.** If rivals match the rate within two quarters, digital human video becomes a commodity input, and differentiation shifts to scripts, voices, and IP — which is where creator budgets should move.

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-aigc-baidu-yijing-video-02-yuan
Free to quote with attribution and a link to the original.
