Vidu Q4 Preview adds voice references and 4K video

Shengshu Technology released Vidu Q4 Preview on October 7, 2026, adding voice references, expanded image references and output options reaching 4K to its video generation lineup. The dated API update notice places the release in preview, with an emphasis on character performance and more complex scenes.

For creators and developers, the useful question is whether those controls make a particular shot easier to finish. That requires looking beyond a polished demonstration. Input restrictions, default settings, queueing and the cost of repeated attempts all matter when a generated clip has to fit an actual production brief.

Two ways to build a shot

The Vidu Q4 product page describes two workflows. Image to video starts with a single picture and a prompt describing motion. Reference to video accepts a collection of images and optional voice samples to guide a scene. Vidu advertises up to 15 image references and three voice references. Those audio samples are conditioning inputs. The page does not promise three separately editable audio tracks.

The reference generation API specifies one to 15 images and up to three MP3 recordings, each lasting three to 12 seconds. It also explicitly excludes the reusable subject invocation method during this preview. An existing integration built around saved subject identities therefore needs to check its input path before treating Q4 as a direct replacement.

Clip length depends on which documentation a reader follows. The product FAQ advertises reference output from one to 16 seconds. The developer reference instead accepts three to 16 seconds, matching the Q4 pricing table. Developers should use the API range when validating requests.

The documented Q4 modes also set the scope of this release. Other tools appearing in Vidu’s navigation, such as text to video, do not by themselves establish that the preview supports those routes. Choose the model specific documentation before designing a workflow around a platform menu.

Quality claims still need a production test

Vidu’s update describes more natural performances, smoother movement, better effects and faster generation. It supplies no measured latency, test protocol or error rate alongside those claims. ByteForward has not run a generation benchmark or independently tested the model. There is no basis here for ranking Q4 against competing systems.

The company announcement also advertises 4K video with 10 bit color. No output file was inspected for this report, and the reviewed material does not establish whether the 4K result is generated at that resolution or enlarged within the service.

A useful trial would use a scene the team already knows how to judge. Check whether the face remains recognizable through a turn, whether dialogue timing fits the performance, and whether a planned camera change preserves the intended action. Record the unsuccessful attempts as well as the strongest take. That makes the result relevant to the job being considered.

API defaults can change the experiment

The image to video API accepts one starting frame in PNG, JPEG or WebP format. It permits an image up to 50 MB but caps the HTTP request body at 20 MB. Inline encoding must fit that smaller request limit. Both API workflows default to a five second clip at 720p, with sound enabled.

This workflow also defaults to using a recommended prompt through the is_rec option. That matters when testing prompt adherence. Explicitly choose whether the service should recommend wording or use the supplied prompt, and record that choice. A silent output can be requested by disabling audio. These controls should be deliberate parts of a comparison.

A November 2025 update lists 10 extra credits for recommended prompts. Confirm Q4’s billing before assuming those prompts are included.

Generation uses an asynchronous task workflow. Keep the returned task identifier, check completion, and retrieve the finished asset. The reference API says creation and cover URLs expire after 24 hours. Applications should save required results rather than use those temporary addresses as permanent media links.

Capacity is another constraint. Vidu’s usage and limits guide sets a standard maximum of five simultaneous tasks per organization. Extra tasks wait in a queue, and capacity can fall under load. The limit applies across the organization rather than separately to each API key. Higher or guaranteed concurrency needs a separate arrangement.

That guide illustrates queueing with tasks that finish in 30 seconds. It is a hypothetical example, not a measured Q4 turnaround time. A batch deadline should be based on observed completion times under the intended workload, including the time requests spend waiting.

Pricing needs a resolution and a clear denominator

The launch announcement quotes USD $0.045 per generated second at 540p and USD $0.39 at 4K. It also announces a 30 percent promotion for image and reference generation through November 30, 2026. The wording does not clarify whether those dollar figures already include the discount. Applying another reduction would produce an unverified price.

The separate API pricing reference listed these charges per generated second for both Q4 workflows. These rates were retrieved from the documentation before our final browser check showed an account information prompt.

• 540p uses 9 credits
• 720p uses 19 credits
• 1080p uses 24 credits
• 2K uses 38 credits
• 4K uses 78 credits

For a ten second clip, that table implies 240 base generation credits at 1080p or 780 base generation credits at 4K. Its published credit unit price is in RMB, at 0.03125 per credit. It should not be relabeled as a dollar rate or silently combined with the launch promotion. Confirm the account’s actual price before committing a budget.

Vidu says failed generations, including moderation rejections, consume no credits. A successful generation that misses the creative brief is a different budgeting concern. Keep a separate count of the outputs an editor would actually use.

Judge the cost of a usable shot

For an initial evaluation, set a fixed brief and a spending ceiling. Keep the reference material, target duration and review criteria consistent. Log generated seconds, total elapsed time and the number of accepted clips. If a shot takes several attempts, divide the total spend by the footage that survives review rather than presenting the cheapest single attempt as the production cost.

This preview is worth investigating when a project needs more explicit guidance over characters, voices and scenes. The decision to adopt it should rest on repeatable results from that project. Start with a small set of representative shots, inspect the exported files and measure the work needed to finish them. Those results will be more useful than an unsupported claim that a new model is universally faster or cheaper.

Featured photograph shows an ARRI Alexa cinema camera as an illustration of filmmaking. Arri Alexa by Sean P. Anderson, via Wikimedia Commons under CC BY 2.0.