An indie animator sketches a character and asks an AI video model to put her in three different scenes. In the first clip she has round glasses; in the second, none; by the third her face is someone else entirely. For most of AI video's short history, that flickering identity has been the genre's tell — the reason "AI video" still reads as a gimmick to producers who need a story to hold together. Vidu, built by Beijing's ShengShu Technology (生数科技), is one of the Chinese models betting that solving consistency, not just prettiness, is the real prize.
What Vidu is
Vidu was the first video-generation model released by a Chinese company, launched in April 2024 and positioned from day one as a competitor to OpenAI's Sora. ShengShu itself was founded in early 2023 by Zhu Jun (朱军), a Tsinghua University alumnus, and grew out of Tsinghua research. Reuters, reporting on the company's 2026 funding, describes ShengShu as the first Chinese firm to ship a video-generation model and notes that Vidu has since iterated through several versions, including the Vidu Q3 release earlier in 2026.
The product today spans two lines. The Q series is the core video model: text-to-video, image-to-video (with optional first and last frames), and a reference-to-video mode, with output up to 1080p. The S series is a streaming, real-time interactive model aimed at digital human (数字人) and live interaction use. This article focuses on the Q series and the consistency problem it targets.
The consistency problem and "reference-to-video"
The headline feature is what Vidu calls multi-subject reference — the ability to upload three or more reference images and fuse their visual elements into one coherent video. The practical win is character and object consistency: instead of hoping a prompt preserves a face, you hand the model the face. A "subject library" lets creators store people, props, and objects so they stay stable across generations.
This matters because consistency is where text-to-video models historically broke down. A single beautiful clip is easy; a sequence where the same person appears in shot after shot, with stable clothing and proportions, is hard. Vidu's reference approach trades some open-ended imagination for reliability, which is the trade professional users actually want.
Beyond the marketing: what it does well
Vidu's own documentation emphasizes three strengths worth taking seriously:
- Large motion dynamics. The company argues its clips avoid the "AI PowerPoint" stiffness — static shots with tiny movements — and instead produce real camera and body motion, which suits effects and stylized scenes.
- Anime stability. Text-to-video and image-to-video both handle animated styles with less drift than many rivals, a reason animators gravitate to it.
- Prompt adherence. Image-to-video follows instructions well enough that iterating on a look ("card drawing," in creator slang) is efficient.
For short-form storytelling — animated shorts, social clips, product videos — those three properties line up with what creators actually ship.
The business behind the model
Consistency alone does not fund a foundation-model lab. In April 2026, ShengShu closed a Series B of roughly RMB 2 billion (about $293 million) led by Alibaba Cloud, with investors including the China Internet Investment Fund and others, according to Reuters. The company said the money would support development of a "general world model" that simulates physical environments and interaction — a step it frames as progress toward general intelligence in the physical world.
That funding context explains the strategy. Video models are expensive to train and serve, and the winners will need both a compelling product and a balance sheet. ShengShu's bet is that owning a defensible consistency feature plus a broader world-model roadmap makes it more than a Sora clone.
Who is actually using it
The consistency angle is not academic. Social-media creators use Vidu to make viral clips — a hug, a photo coming to life — where a stable face is the whole point. Animation and game teams generate storyboard-to-final clips from scripts. E-commerce teams animate a static product image into a rotating, expressive hero shot. Short-form animated content, the Chinese 动漫短剧 wave included, is a natural fit because stylized characters tolerate the model's remaining physics weaknesses better than photoreal humans do. The common thread is volume: these users need many clips fast, with a character that survives from one to the next.
Honest limitations
Vidu's marketing claims about benchmark leadership come from the company and its partners, not from independent head-to-head testing we can verify. Real-time and long-form quality on the S series is a separate question from the Q series' clip generation, and the two should not be conflated. Like every text-to-video system, Vidu still struggles with precise physics, legible text in frames, and complex multi-character choreography. And the "consistency" win is relative: reference images help, but sustained, minute-long narrative coherence remains out of reach. Treat vendor "ranked first" claims as unverified until an independent lab republishes them.
What readers can do now
- Test the reference-to-video mode with your own character sheet. Upload three angles of a mascot or actor and generate a short clip to see how stable the identity actually stays.
- Use it for shots, not whole films. Build a scene from generated clips and edit them together; the model is strongest as a shot factory, weakest as a director.
- Separate the Q and S series in your evaluation. If you need interactive digital humans, judge the S series on its own latency and interaction metrics rather than the Q series' clip quality.
