Vidu Q4 API

Vidu Q4 is the newest video model from ShengShu Technology, the Beijing company behind the Vidu brand, and it is built around one idea: AI characters should be able to act. ShengShu opened it to the public on 7 October 2026 as Vidu Q4 Preview, an early look at its next flagship, and pitched it on performance rather than spectacle. The model coordinates facial expression, emotion, body movement and voice in a single pass, so a character who is supposed to be nervous looks nervous before saying a word.

The feature list backs the pitch. Q4 accepts up to 15 reference images to lock down characters, costumes, props and locations, and up to three short voice recordings to pin down how each character sounds. It renders at up to 4K, with 10-bit colour at 2K and 4K, and generates synchronised dialogue and sound effects alongside the picture. Clips run up to 16 seconds and can contain several shots.

It also arrived with a strong first score. A day after launch it sat third on the Artificial Analysis image-to-video arena, 123 Elo points above its predecessor. On Unicorn API it comes as two models, Vidu Q4 Image To Video and Vidu Q4 Reference To Video, each with all five resolution tiers from 540p to 4K.

ShengShu and the road from Vidu Q3

ShengShu grew out of research at Tsinghua University, with machine learning professor Zhu Jun as chief scientist. Its first Vidu model in 2024 was built on U-ViT, the lab's own architecture that combines diffusion with a transformer backbone, and ShengShu has offered Vidu through an enterprise API since version 1.5. The company sells Vidu both as a web app for creators and as an API platform for studios and developers.

The direct predecessor is Vidu Q3, released at the end of January 2026. Q3's headline was native audio: it generated dialogue, sound effects and music together with the picture, in one pass, for up to 16 seconds at 1080p and 24 frames per second. In April ShengShu added Q3 Reference-to-Video, which let creators feed in subjects, environments, costumes and props and keep them consistent across a clip, with multilingual dialogue.

Q4 keeps that foundation and shifts the emphasis. Length stays at 16 seconds and audio stays native. What changes is how much control the creator has and how good the characters look while they perform.

Vidu Q4 is a closed model. ShengShu has not published weights, a parameter count or a technical report for it, and it is available only through Vidu's own products and API partners such as Unicorn.

What changed in Vidu Q4

Acting, not just moving

ShengShu's main claim is "more nuanced expressions, clearer emotions, and more natural movement". The model is designed to tie the face, the body and the voice together, so that a line of dialogue lands with matching eyes, posture and timing. One early creator noted that even in a pause with no dialogue, the characters' eyes and expressions alone carried the emotion. That is the kind of moment earlier video models tended to fill with blank stares or random fidgeting.

Voice references

The biggest new input is audio. Reference-to-video accepts up to three MP3 clips of 3 to 12 seconds each, and the prompt can assign them to characters ("the woman in reference image 1 says … using reference audio 1"). The voice samples guide identity, emotional delivery and lip sync, which means a recurring character in a series can keep the same voice from episode to episode instead of getting a new one with every generation.

Fifteen image references

Q4 combines between one and fifteen reference images in a single generation. They can be faces, wardrobe, products, props or whole environments. For short dramas, ads and product films this is the difference between "a woman in a red coat" and this woman in this coat in this shop.

4K with 10-bit colour

Q3 stopped at 1080p. Q4 adds 2K and 4K, and both higher tiers carry 10-bit colour, roughly a billion shades instead of the 16.7 million of 8-bit video. ShengShu pitches this at professional finishing: smoother gradients in skies and skin, and more room to grade the footage after export.

Action, camera and effects

ShengShu also reworked fast motion. It reports more coherent cuts and more dynamic camera moves in fights and chases, fewer deformed limbs in martial arts and running, and better blending of explosions, fireworks, smoke and particles into the scene. Beta testers described steadier camera movement and shot transitions and better adherence to instructions.

How Vidu Q4 scores on the image-to-video arena

Artificial Analysis runs blind comparisons in which viewers pick the better of two clips made from the same starting image. On 8 October 2026, Vidu Q4 Preview debuted third on its image-to-video board with an Elo of 1,179 (plus or minus 10, from 5,543 samples). Only MiniMax H3 Max (1,195) and MiniMax H3 (1,181) were ahead of it, with Google's Gemini Omni Flash one point behind.

Two caveats keep this in proportion. The top six models on that board sit within overlapping confidence intervals of about ten points, so "third" is effectively a tie with the leaders rather than a clear ranking. And the vote count is still young for a model that is days old.

The jump from its predecessor is not ambiguous, though. Vidu Q3 Pro sits 19th on the same board at 1,056, so Q4 gained 123 points, more than the entire spread across the top six places. Q4 does not appear on the text-to-video board at all, because it has no text-to-video mode.

Specs and the two Vidu Q4 models

On Unicorn both models default to 5 seconds at 720p with audio on. Images can be JPG, PNG or WEBP; the first-frame image can be up to 50 MB. Setting audio to false gives a silent clip.

The image-to-video model treats your picture as the opening frame and animates it according to the prompt, or with its own judgement if the prompt is empty. The reference model builds a new scene from your images and voices, and your prompt refers to them by position: "reference image 2", "reference audio 1".

Where Vidu Q4 shines

Dialogue scenes and short drama. This is the use case Q4 was designed around. Two characters, a few lines each, a reaction shot: ShengShu says the model handles eye lines, micro-expressions and lip movement better than earlier Vidu releases, and the voice references keep the cast sounding consistent.

Recurring characters and products. With up to 15 images, you can hand the model a character sheet, an outfit, a product and a location, and get a clip where all four stay recognisable. Ad and e-commerce teams can keep a product's shape and label intact while changing the setting.

Action with effects. Fights, chases, sparks and smoke are a stated focus, and Q4's multi-shot clips can cut between angles within one generation instead of needing separate clips stitched together.

Finishing quality. If a clip has to survive a big screen or a colour grade, 4K with 10-bit colour is a real step up from 1080p. One early tester also singled out Q4's rendering of Chinese text on screen.

Long, detailed prompts. The 20,000-character prompt limit leaves room for a full shot list with blocking, camera moves and dialogue.

Limits to know before you build on it

It is a preview. ShengShu describes Q4 Preview as an early release to gather feedback before the full Q4, and says behaviour and details may change before then.

No text-to-video. Every generation starts from at least one image. If you only have a script, you need to make a still first, for example with an image model, or pick a model that generates from text.

Sixteen seconds is the ceiling. Q4 does not extend clip length over Q3. Longer scenes have to be assembled from several generations, and keeping continuity across them is your job, helped by the reference images.

Higher resolution is not higher accuracy. 4K sharpens what the model draws; it does not guarantee a consistent face, correctly spelled text or perfect lip sync. Check pronunciation, timing and mouth shapes before publishing anything with dialogue.

Little is published. There is no technical report, no parameter count and no list of supported dialogue languages for Q4, and most quality claims so far come from ShengShu and early partners rather than independent testing.

Prompting tips for Vidu Q4

Write like a director. Q4 responds to blocking, emotion and camera language, so describe what the character feels and how the shot moves, not only what is in frame. Put dialogue in quotation marks and say who speaks it. For reference-to-video, name every input by its number and give each a role.

An image-to-video prompt for a portrait still:

The woman slowly lifts her eyes from the letter, swallows, and forces a small smile. She says quietly, "He's not coming back, is he?" Slow push-in to a close-up, soft window light, faint rain on the glass.

A reference-to-video prompt with two characters and a location:

The chef in reference image 1 and the food critic in reference image 2 sit at the counter of the diner in reference image 3. The critic takes one bite, pauses, and raises an eyebrow. The chef leans in and says, "Well?" using reference audio 1. Cut to a close-up of the critic grinning. Warm evening light, handheld camera.

An action shot:

The swordswoman in reference image 1 sprints across the rooftop, vaults a chimney and lands in a crouch as fireworks burst behind her. Tracking shot from the side, then a low-angle cut as she rises. Sparks drift through the frame.

Start at 540p or 720p while you iterate on blocking and timing, then render the keeper at 1080p, 2K or 4K. Lock the reference images and voice clips before the final render, since changing them alters the character, not just the detail. Keep audio on while drafting dialogue scenes, so you can judge lip sync and timing at the same time as the blocking.

Vidu Q4 against other video models on Unicorn

MiniMax H3. The model just above Q4 on the image-to-video arena. MiniMax H3 Reference to Video is the closest alternative for reference-driven work, and H3 also offers text-to-video, which Q4 lacks. Pick Q4 when you need voice references, 15 image slots or 4K output.

Kling 3.0 Omni. Kling 3.0 Omni is Kuaishou's Omni model and a natural comparison for teams already working in the Kling ecosystem. Q4's edge is the explicit audio references for consistent voices.

Seedance 2.5 and Wan 3.0. ByteDance's Seedance 2.5 and Alibaba's Wan 3.0 are the other flagship-class options in the catalog and worth testing side by side on the same brief, especially for scenes without much dialogue.

HappyHorse 1.1. Alibaba's HappyHorse 1.1 Reference To Video is another multi-reference model if you want a second opinion on a character-consistency job.

The short version: choose Vidu Q4 when performance is the point, meaning characters who talk, react and stay recognisable across shots, or when the output needs 4K. Choose a text-to-video model when you are starting from words alone.

Getting started with Vidu Q4

The quickest way to try it is the Playground: open Vidu Q4 Image To Video in the studio, drop in a still and describe the scene, or switch to Vidu Q4 Reference To Video to combine several images and voice clips.

For code, create a key on the API keys page and submit a job to /v1/jobs with model set to vidu-q4-image-to-video or vidu-q4-reference-to-video, plus image_url or reference_image_urls, a prompt, and optionally duration, resolution, audio, aspect_ratio and reference_audio_urls. The API tab on this page has full request examples, and the docs explain how jobs work.

Verdict

Vidu Q4 is ShengShu's clearest statement yet that the next frontier in AI video is performance. Voice references, fifteen image slots and 4K 10-bit output make it one of the most controllable reference-driven models available, and its third-place debut on the image-to-video arena, 123 points above Q3 Pro, shows the gains are visible to viewers, not just on a spec sheet. It is still a preview, it cannot start from text alone and it stops at 16 seconds. But for dialogue scenes, recurring characters and short-form drama, it is the model to test first.