NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

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.

2026-11-18 · 925 words · NeuroAI
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 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.

Related coverage

Cite this article

NeuroAI (脑机网), "Baidu's YiJing Pushes Digital Human Video to US$0.03 a Second — and Adds Micro-Expressions to Kill the 'AI Face," 18 November 2026, https://neuroai.site/a/na-aigc-baidu-yijing-video-02-yuan

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Generative Media” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories