Baidu has listed ERNIE-Wizshot-ME on Qianfan, its cloud AI platform, offering short-video creators and small production teams a service for scripted scenes. The model combines character images, dialogue and reference voices to generate short talking-character videos.
Jiemian reports that Baidu unveiled the model on November 12, 2026, with API access for enterprise customers and partners. Its task-creation documentation carries a November 11 update date; the exact start of availability remains unclear.
Baidu advertises 4–15-second, 1080p clips for close-up drama, storytelling and presenter-style videos. Its claims of natural, expressive performances are vendor claims, not independently established quality comparisons.
What creators need to prepare
According to Baidu’s creation guide, each character needs an image and a 10–60-second WAV recording containing speech. That recording is a voice reference, not the output video’s length. Requests also need images for individual shots, camera directions, speakers, dialogue and emotion instructions. Optional controls cover performance and voice delivery.
The practical implication is that creators still need to prepare the cast, voices and scene before generation. This is a direct video service, distinct from ERNIE 5.5 Flash Preview, Baidu’s text-only model whose creative demonstrations use external tools.
The documented access route is an API—a connection for software to use the model. Qianfan’s setup guide calls for console sign-in and an API key configured for the relevant resources. It does not establish worldwide account eligibility, and the listing should not be read as a ready-made editing app.
What the list price buys
Baidu lists CNY 0.70 per generated second, or 3.50 Chinese yuan for five seconds. That is a generation charge, not the cost of a finished production.
Illustrative calculation: six separate 10-second clips would cost CNY 42 for one generation each: 6 × 10 × CNY 0.70. Producing three versions of every clip would cost CNY 126, assuming all 18 generations succeed and are billed at list price. Asset preparation and editing are additional; discounts, taxes and failed-task charges are not included in this estimate.
Limits that affect a production schedule
The API documentation checks that a scene’s generated dialogue plus silence falls within 4–16 seconds, while the product page advertises 4–15 seconds. That boundary is unresolved: do not plan around a guaranteed 16-second output. Scripts outside the API’s permitted duration fail.
The model catalogue lists one running task and one queued task. This constrains how much work can be submitted at once; it does not tell a team how many clips it can finish per hour. Rendering speed remains unverified.
Generation is asynchronous: submit a task, then check its status for completion or failure. Successful video links expire after 24 hours, although Baidu says the content is retained for at most 30 days. Teams should save completed clips to their own storage before the link expires.
For a team considering a scripted close-up or presenter clip, the immediate decision is whether its prepared assets and short script fit the service. A low per-second rate alone cannot establish the budget or deadline for a longer finished video.