You shot a 30-second ad and hated the lead actor's jacket. In a normal production that means a reshoot, a new wardrobe call, and a day lost. In a generative video tool, it should mean changing one reference and regenerating the shot.
That gap — between "regenerate everything" and "change one thing" — is the gap ShengShu Technology (生数科技) is trying to close with Vidu Q4 Preview, the first preview build of its new flagship video large model (大模型), released on October 8, 2026.
What actually changed in Q4 Preview
The headline upgrades are about control, not just prettier frames:
- Reference capacity: up to 15 reference images and 3 reference audio tracks in a single generation. The model uses them to lock a character's face, costume, props and voice across a scene.
- Resolution: 2K and 4K output, aimed at broadcast and commercial specs rather than social clips.
- Performance focus: the build emphasizes character acting, camera language and complex visual effects — the parts that decide whether a clip reads as "film" or "demo."
ShengShu's co-founder and CEO Luo Yihang framed the release around lowering the barrier: "Artificial intelligence only becomes sustainable productivity when it enters real scenarios and solves real problems. We want model efficiency to keep turning into lower thresholds, so more individuals, teams and enterprises can afford to use it."
The number that matters: 0.09 yuan per second
Flagship video generation has usually been priced in chunks that scare off small teams. Vidu Q4 Preview opens at a promotional floor of about RMB 0.09 per second (≈ US$0.013, or roughly US$0.76 per minute of footage).
At comparable spec and billing, ShengShu says the same budget now produces up to five times the video volume it did before. Whether a creator reaches exactly 5x depends on length, resolution and how many reference assets they feed in — but the direction is clear: the cost of trying one more version is falling fast.
That matters most in places where version count is the product. In microdrama (短剧) and anime production, teams pre-visualize shots, test character blocking and validate camera rhythm before a single real frame is shot. In e-commerce and advertising, brands spin dozens of creative variants for different channels and audiences. When each variant costs pennies instead of a production day, the creative loop tightens.
Why reference control is the real story
For two years, the practical ceiling on a single generative video sat between 8 and 15 seconds. Longer clips meant stitching fragments, and stitching is where faces drift, lighting shifts and continuity breaks.
Vidu's path has been different. The team proposed the U-ViT architecture in 2022, unveiled Vidu in April 2024, then pushed reference-to-video, multi-subject consistency, synced audio, multi-shot narration and real-time interaction. Q4 Preview concentrates those threads on acting and camera work, using multiple images and audio to hold creative elements steady.
According to Li An, chief scientist at BrainNet (脑机网), the real shift in Chinese generative video this year is less about longer clips and more about giving creators reliable control over a character's face, voice and wardrobe across a whole scene.
From showcase to production line
When video is billed per finished second, the unit of creative work flips. A director stops treating each generated clip as a precious, expensive artifact and starts treating it as a draft. That is the same psychological shift photography saw when film gave way to digital memory cards: you take more, you keep the best.
For microdrama (短剧) studios especially, the math is brutal and therefore instructive. Traditional live-action shorts can run into the hundreds of thousands of yuan to produce; AI pipelines have pulled that down sharply, and per-second pricing pushes the marginal cost of "one more take" toward zero. Vidu's 5x budget claim is really a claim about how many takes a team can now afford.
Where it fits in the Chinese video race
Vidu Q4 Preview lands in a crowded, fast-moving field. In the summer of 2026, ByteDance's Seedance 2.5, MiniMax H3 and Alibaba's Wan 3.0 all shipped major video updates. Each pushed clip length, native audio or document-to-video workflows. Vidu's angle is reference density and acting control — a bet that predictable, editable characters beat raw spectacle for production teams.
ShengShu says it is already working with film, content-platform and cloud partners on generative audiovisual creation, microdrama (短剧) production and enterprise API access.
Honest limitations
A few things the launch does not settle:
- "Preview" means preview. This is the first preview build of a new flagship, not a finished, fully documented product. Expect capability and pricing to move before general release.
- Reference counts are ceilings, not promises. Hitting 15 images and 3 audio in one stable generation depends on prompt quality, asset preparation and the specific shot. Real output varies.
- 4K claims need the full spec sheet. Resolution support was stated by the company and corroborated by Beijing municipal science authorities; final bitrate, motion coherence at 4K, and per-second cost at 4K are not fully broken out yet.
- Independent benchmarks are thin. Most performance framing comes from the vendor and Chinese state-linked media. Outside, head-to-head testing against Seedance 2.5 or Wan 3.0 is not yet public.
What readers can do now
- If you make short dramas or ads, run a 15-reference test: build one character from a face, a costume shot and a voice clip, then regenerate only the wardrobe to see how stable the identity stays.
- Benchmark your own cost: take a real 30-second brief and price it at the 0.09-yuan floor versus a traditional reshoot; the gap is the business case.
- Wait for the spec sheet before committing pipeline: if you plan to depend on 4K and multi-audio, hold until ShengShu publishes full resolution and billing details.