Vidu Q4 Preview API Guide: Reference Images, Audio and Task Status

Vidu Q4 Preview API requests can combine reference images and audio to guide a generated video. The practical workflow has two parts: create a request that expresses a consistent scene, then track the asynchronous task until a usable output is returned. A task ID by itself is not a finished clip.

The official reference-to-video documentation provides the reviewed limits and task endpoints. This guide explains a pilot workflow; it does not report a Vidu generation we ran or a measured success rate.

Vidu Q4 Preview API limits and input choices

The documented model identifier is viduq4-preview, and the create endpoint is https://api.vidu.com/ent/v2/reference2video. Keep the selected regional platform and API host together; credentials or documentation from another host should not be mixed casually into the integration.

The API page also gives a 20,000-character prompt limit and a 20MB request-body limit. Large Base64 inputs can increase the body size, so check the encoded request rather than only the original image's file size.

A preview-specific note says subject invocation is not currently supported, even though the broader page contains subject-style examples. Use the direct image references for the reviewed preview path and avoid assuming named subject-library syntax works.

Prepare consistent image and audio references

Select images that show the intended subject clearly. A mixture of different faces, object shapes or lighting conditions can make a consistency test ambiguous: the references themselves may disagree about the desired result.

Begin with one subject and one action. For example, a single object turning slowly on a table is easier to inspect than several characters performing different actions across many camera cuts. Add complexity only after the basic reference workflow produces an acceptable clip.

Give each audio reference a clear role and confirm it meets the documented format and duration. Reference audio and the audio output flag are distinct controls. Supplying a sound reference is not the same as deciding whether the returned clip should use audiovisual output.

Keep input rights and the intended use attached to the project record. For a recurring creative workflow, knowing which reference files were approved is as useful as knowing which prompt produced the clip.

Submit one reference-video task

Prepare the request body in a file so that inputs can be reviewed and reused. The following is a body template; replace the deliberately marked image URL with an accessible reference you are authorized to use:

{
  "model": "viduq4-preview",
  "prompt": "The referenced ceramic cup rotates slowly on a plain table. Keep the camera fixed and the cup shape consistent.",
  "images": ["REPLACE_WITH_AUTHORIZED_ACCESSIBLE_IMAGE_URL"],
  "sounds": [],
  "audio": false,
  "duration": 5,
  "aspect_ratio": "16:9",
  "resolution": "720p"
}

Save the completed body as request.json. The create example in the reviewed English endpoint reference uses a Token authorization prefix:

curl https://api.vidu.com/ent/v2/reference2video \
  -H "Authorization: Token $VIDU_API_KEY" \
  -H 'Content-Type: application/json' \
  --data-binary @request.json

The same page's query example uses Bearer, and regional documentation can differ. These snippets follow the cited English examples; verify the authentication contract for your account and host before embedding it in a reusable client. Do not diagnose a header mismatch as a creative-model failure.

Save the returned task ID and response status. Do not automatically resubmit on every delay: duplicate create requests can start additional generation jobs. First determine whether the initial task exists and is still processing.

Track completion and save the output

The documented query path is /ent/v2/tasks/{id}/creations. For the returned task ID, its reviewed example is:

curl "https://api.vidu.com/ent/v2/tasks/$VIDU_TASK_ID/creations" \
  -H "Authorization: Bearer $VIDU_API_KEY"

Keep task creation separate from successful output retrieval. Record the returned state, any error code and the creation URLs. A successful HTTP query can report a task that has not yet succeeded.

If using callbacks, follow Vidu's documented signature verification. The page describes processing, success and failed notifications, with retries for terminal-state callback delivery. A callback handler should cope with repeated delivery rather than creating another job each time it receives an update.

Once the task succeeds, inspect and store the result according to the returned URL conditions. The reference notes a limited validity period for watermarked creation URLs; a link in a task record is not a durable media archive.

Watch the clip itself. Check whether the subject remains recognizable, the requested action completes, camera movement follows the brief and audio behavior matches the request. An output file existing is a workflow milestone, not proof of creative acceptance.

Compare usable video costs and preview behavior

Record charged credits or usage from the account and task result. Compare settings using the same brief and references, including the extra attempts needed to obtain an accepted clip. Do not convert a rate from another model version into a Q4 price without checking the applicable schedule.

Test higher resolution only after the basic scene works. Keep duration and references fixed while comparing it, so that a billing or quality difference has a clear cause. Preview behavior and supported limits can change, making the exact model name and review date useful parts of the record.

For a related comparison, our Kling API availability guide addresses a different video route. Our MiniMax H3 prompt guide provides another prompting context. Preserve each provider's own request schema when testing the same creative job.

Conclusion: verify the full Vidu Q4 task lifecycle

Vidu Q4 Preview API evaluation should begin with consistent references and a bounded scene. Confirm the request contract, save the task ID and wait for a successful creation result before reviewing the clip. Compare actual charges across accepted outputs, with preview restrictions and authentication differences kept explicit.

Frequently asked questions

Does a task ID mean the clip is ready?

No. Query the task or process a verified callback, then inspect the successful creation result and the clip.

Can I use named subject-library invocation with this preview?

The reviewed reference explicitly says the preview does not currently support subject invocation. Use the supported direct-reference path.

Do higher resolution and more references guarantee consistency?

No. Coherent inputs, a clear scene and output review remain necessary. Evaluate the exact settings rather than treating larger limits as a quality guarantee.