Vidu Q3 Split Reference-to-Video Into Three Products. That Is a Packaging Decision as Much as a Model One.

Most video model releases get announced once and then show up everywhere under a single name. Vidu did the opposite with its Q3 reference-to-video line.

On September 14, Alibaba Cloud's Model Studio listed three separate Vidu Q3 reference-to-video models. The first, viduq3-mix, is a general-purpose model that takes reference images and a text prompt and blends the subjects from those images into the described scene. The second, viduq3-ad, is built for advertising, with marketing-grade scene cuts, camera movement, and direct audio output; you upload product images and get an ad video. The third, viduq3-drama, targets drama and AI manga production, with character consistency, motion effects, and emotional expression tuned for story-driven content.

All three run at $0.14 per second for 720p or 1080p, with a rate limit of 5 requests per second. The pricing is identical across the three, which tells you the differentiation is not in cost but in output behavior.

The problem Vidu actually set out to solve

Vidu comes from Shengshu Technology in Beijing. Its Q3 model launched on January 30, 2026, and immediately landed on the benchmark circuit. Artificial Analysis ranked it first in China and second globally at launch, ahead of Runway Gen-4.5, Google Veo 3.1, and OpenAI's Sora 2, on a score of 1241.

The three technical shifts behind that ranking are worth naming. Q3 was among the first long-form video models to produce synchronized dialogue, sound effects, and background music in the same pass as the visuals, rather than bolting audio on in post. It extended single-run clips to 16 seconds, which was the longest among major models at the time. And it built a reference-to-video system where users upload three to seven images of a character, object, or style and the model uses those as visual anchors in later generations.

The reference system is the interesting one. Most video models in 2026 compete on the same axis: make a single clip look impressive. Vidu's bet was different. The pitch was never "look at this one gorgeous shot." It was "look at this same character across ten shots and notice they still look like the same person."

That is a harder problem, and it is the one series content actually depends on. In a comparison against Runway Gen-4, Kling 3.0, Luma Ray, Pika 2.0, and Sora 2 using the same character prompt across six scene variations, Vidu Q3 maintained facial and outfit consistency in five of the six tests.

Consistency is what separates a tool for one-off clips from a tool for a series. For anime creators, short-drama producers, and anyone building brand content around a recurring character, it changes what is possible.

Why three SKUs matters

Splitting one model family into three named products looks like a marketing choice. It is closer to a procurement decision.

An advertising team and a short-drama studio do not buy the same thing, even if the underlying transformer is similar. The ad buyer needs product fidelity, clean scene transitions, and audio that can sit under a brand voice. The drama buyer needs a character to hold identity across episodes and shots, with expressions that read as performance. Packaging those as separate model IDs with separate documentation lets each buyer evaluate and budget for exactly one job, rather than reading a general-purpose feature list and guessing.

The general-purpose mix model then serves everyone outside both boxes, a reasonable structure for a market where the buyer is increasingly a production team with a deadline rather than a hobbyist testing prompts.

Vidu Claw and the pipeline layer

Packaging the models is half the strategy. The other half is an assembly layer. Vidu Claw is an AI creation automation engine built on the open-source OpenClaw framework and powered by Q3. You connect your Vidu tokens, define Skills, and automate the path from idea to finished video. A single prompt like "create a short-form ad for a running shoe targeting young women on Instagram" is enough for Vidu Claw to generate a concept, script, storyboard, visuals, voiceover, music, and a final cut.

Independent reviews have been mixed in a useful way. A reviewer testing it on a product commercial found the platform genuinely free and accessible for students and beginners, but reported that it struggled with precise structural detail. In one case it kept rendering a detail in a way that stood out even after repeated prompt corrections, and the output carried a noticeable AI flavor, with lighting and color that felt cheap.

That is the honest trade-off. Vidu Claw is a productivity tool for solo creators and small marketing teams that need to move from concept to a usable draft quickly, rather than a replacement for a professional production team. When it works, it compresses a five-day advertising production cycle into a day. When it does not, the quality issues are hard to overlook.

The ecosystem Vidu is building around

The reference-to-video focus connects to a distribution story. In February 2026, the software company Wondershare announced a strategic investment in Shengshu, with the two planning to work on AI manga. The Vidu Q3 model was integrated into a full-chain AI manga production platform used to keep characters, voices, and scenes consistent across episodes. Alibaba Cloud's Bailian platform also added Vidu Q3 as a third-party model in May 2026.

So the model travels through at least three channels: direct web and API access, a creation platform aimed at manga and short drama, and a cloud model marketplace. Each channel has a different buyer with a different definition of "good enough."

Where the money actually is

The shift toward series content is not a creative preference. It is where the budgets are. A short drama is a multi-episode product with a recurring cast, and its economics depend on producing far more footage than a single ad or a one-off clip. If a model forces the creator back to zero on character appearance for every shot, the savings from generated footage get eaten by consistency fixes. Vidu's bet is that the team willing to pay $0.14 per second is not buying a single beautiful shot but a way to keep the same person recognizable across an entire run.

That is also why the reference system matters more than the raw quality score. A model that ranks second globally on a benchmark is impressive, but the benchmark measures clips. The reference feature measures series. The two are not the same purchase, and Vidu is selling to the second one.

Where it does not yet deliver

Reference consistency is not perfect, and five out of six is not six out of six. Where the sixth case failed, the tool delivered a face and outfit that drifted from the source, which is exactly the failure that breaks a series. Cost is real at $0.14 per second, which adds up quickly across a 16-second clip and becomes significant at series scale, where hundreds of shots are in play. The pipeline layer produces drafts rather than finished commercials, and reviewers have found it cannot reliably fix a structural detail even with repeated prompt corrections.

There is also a structural limit worth naming. Reference-to-video anchors identity, but it does not yet solve continuity for a whole production. Someone still has to make the directorial choices that turn a set of consistent clips into a story, and the model has no opinion about pacing, coverage, or whether a scene should exist at all.

What to watch

The three-SKU structure is the part other vendors are likely to copy. If reference-to-video becomes the default way series content gets made, expect more model families to ship job-specific variants instead of a single flagship. The questions that matter next are whether the consistency holds at longer durations and harder camera work, and whether the pipeline layer gets good enough to sell to a studio rather than a solo creator.