Vidu Q4 Preview: 15 References, 3 Voices, Real Acting

Most AI video can make a face move. Very little of it can make a face act. The eyes stay a beat behind the line, the voice belongs to somebody else, and the emotion arrives as a stock expression pasted on top. That's fine for a product spin. It falls apart the second you try to tell a story with a character in it.

ShengShu thinks it has a fix. On October 7, 2026 it released Vidu Q4 as a public preview, calling it the first look at its "next-generation flagship model for expressive audio and video." The pitch is not more pixels, although there are more of those too. It's performance: expression, emotion, body movement and voice moving together. Here's what's confirmed, what's still open, and how to run it on each::labs today.

What Vidu Q4 Preview Actually Is

Vidu Q4 is the newest model in ShengShu's Vidu video family, and the Preview label matters. ShengShu shipped it ahead of the full Q4 so creators can use it on real projects and shape the final version. According to the company's launch release, it's available now through Vidu's own web product and API platform, and ShengShu adds that final resolutions, features and terms may still vary by plan and region.

The official Vidu Q4 page lists two modes: image-to-video and reference-to-video. Clips run up to 16 seconds, with image-to-video between 3 and 16 seconds and reference-to-video between 1 and 16. Output goes up to 4K, and the launch release says the 2K and 4K tiers come with 10-bit color. The full ladder named in the release is 540p, 720p, 1080p, 2K and 4K.

Who is it for? ShengShu names independent creators, small studios, short drama companies, ad agencies and film production teams. That list tells you the target better than any spec. This is a model for people who put characters on screen and need them to hold up for more than one shot.

Fifteen References: Build the Whole Cast Up Front

The number everyone will quote is fifteen. Vidu Q4 accepts up to 15 reference images in a single creative setup, and ShengShu spells out what they're for: characters, wardrobe, props, products and environments. One image for the lead's face, one for the jacket, one for the car, one for the street, and the model is asked to keep all of them straight.

That's a real shift in how you work. With most video models you fight for consistency after the fact: generate, notice the jacket changed color, regenerate, pray. A deep reference stack moves that fight to the start, where it belongs. You assemble the cast and the set once, then direct inside it. If you've ever lost an afternoon to character drift across shots, you know why this is the feature that matters.

The Vidu Q4 page also notes that output keeps your source image's aspect ratio. Small detail, big relief: a vertical reference stays a vertical clip, so you're not cropping a short-form edit out of a landscape render.

Three Voices and a Face That Means It

Images define who a character is. Audio references define how they sound. Vidu Q4 takes up to 3 reference audio clips, which ShengShu says guide voice consistency and emotional delivery. The product page goes further, describing voice cloning "with natural emotion and lip-synced delivery," with audio and visuals in sync.

Three voices is the interesting number. One voice is a narrator. Two is a conversation. Three is a scene. That puts dialogue between characters on the table, not just a talking head reading a script, and it's why short drama studios are on ShengShu's list.

Then there's the acting itself. The launch release promises "more nuanced expressions, clearer emotions, and more natural movement." One creator quoted in ShengShu's release put it more plainly: the characters' eyes and expressions alone were enough to carry the emotion. That's a single testimonial, not a benchmark, so treat it as a claim to test. But it points at exactly the gap described at the top of this post. If Q4 closes it, the voice clone and the face finally belong to the same person.

Action, Cuts and Effects That Stay in the Scene

The second half of the pitch is the hard stuff: chases, fights and effects. ShengShu says camera work in Vidu Q4 is designed to follow the action more coherently, with "more coherent cuts and dynamic camera movements in complex action sequences." The product page claims fewer deformations, stutters and artifacts in martial arts, running and combat, which are exactly the shots that usually melt.

Effects get the same treatment. Explosions, fireworks, smoke and particles are meant to fit into the surrounding scene rather than float on top of it. A creator in the release describes explosion light that illuminates the surrounding buildings. That's the tell of a model that understands the effect as part of the world, not a sticker. The same creator also singles out strong Chinese text rendering, a nice bonus for anyone localizing for that market.

Multi-shot is part of the story too. The Q4 page talks about up to 16 seconds of multi-shot storytelling, so a single generation can hold a cut or two instead of one continuous take.

Vidu Q4 vs Vidu Q3: What Changed

ShengShu hasn't published a head-to-head with its previous model, so here's what the official pages support. Vidu Q3 already generated native audio and video in one pass, including dialogue, voiceover, sound effects and music, with a single generation running up to 16 seconds and detailed control over camera language and rhythm, according to Vidu's own Q3 page.

So Vidu Q4 isn't about longer clips. The 16-second ceiling is the same. The upgrades are in control and fidelity: up to 15 image references, up to 3 voice references for consistent cloned voices, 2K and 4K output with 10-bit color, and the performance and action improvements above. Q3 made a scene with sound. Q4 is trying to make the scene with your cast, your voices and your props, and make it look like they're acting.

If you've been building on earlier Vidu models, the shape of the workflow will feel familiar. Vidu Q1 Reference to Video and Vidu 2.0 Reference to Video are live on each::labs right now, and our Vidu Q2 write-up covers how the family's reference approach evolved.

Where Vidu Q4 Fits Next to Seedance, Kling and Wan

October is crowded. Seedance 2.5 and Wan 3.0 both run to 30 seconds with audio and multiple references, and both are live on each::labs. Kling 4.0 is promised for this month with 30-second shots and keyframes, as we covered in what we know about Kling 4.0.

Against that field, Vidu Q4 isn't trying to win on length. Its bet is narrower and, honestly, more interesting: the most convincing acting in a short clip. Voice references for multiple characters, a deep reference stack and a focus on expression put it closest to the short drama and character-ad end of the market. If your clip is a product rotating on a plinth, you don't need any of that. If your clip is two people arguing in a kitchen, you might need all of it.

How to Run Vidu Q4 on each::labs

Vidu Q4 Preview is live on each::labs in two modes: Vidu Q4 Preview Reference to Video for multi-reference scenes, and Vidu Q4 Preview Image to Video when a single frame should set the shot. Either way, the output is only as good as what you feed it. Build the inputs first.

Start with the reference library. Fifteen slots is only useful if you have fifteen good images. Shoot or generate clean, well-lit references for each character, outfit, prop and location, and give every subject its own image. A model like Nano Banana 2.1 can produce consistent character sheets you'll reuse across every clip.

Record your voices next. Short, clean audio of each character, with the emotional range you want them to have. A flat read in, a flat performance out.

Then write prompts like a director, not a describer. Name who speaks, what they feel and where the camera goes:

Two-shot in a small night kitchen. The older sister (reference 1) leans on the counter, tired and trying to stay calm; the younger brother (reference 2) paces, angry. She speaks first in voice 1, quietly. He cuts her off in voice 2. Slow push-in on her face as she goes silent. Warm practical light from the stove, rain on the window.

Run that structure through Vidu Q4 Preview Reference to Video, then try the same prompt on Vidu 2.0 Reference to Video or any model in the image-to-video catalog to see how far the family has come. On each::labs, switching models is a model id change, not a new integration.

Frequently Asked Questions

When did Vidu Q4 come out?

ShengShu released Vidu Q4 Preview on October 7, 2026, through Vidu's web product and API platform. The full Q4 has no announced date yet, and the company says details may change before it.

How many references does Vidu Q4 support?

Up to 15 reference images for characters, wardrobe, props, products and environments, plus up to 3 audio references that guide a character's voice and emotional delivery.

How long are Vidu Q4 videos?

Sixteen seconds at most. Image-to-video runs 3 to 16 seconds and reference-to-video 1 to 16, with multi-shot storytelling inside that window and output up to 4K.

Is Vidu Q4 available on each::labs?

Yes, as of the Preview. You can run it as reference-to-video or image-to-video, and earlier Vidu models like Vidu 2.0 and Vidu Q1 sit right next to it for side-by-side tests.