What Vidu Q3 is, in the publisher's own words
Shengshu Technology describes this model on its own product page under the line "Vidu Q3,为剧而生" — Vidu Q3, born for drama. The page is built around three claims. On audio and picture together it says "音频与画面同步输出——对话、旁白、音效、音乐四类音轨同步生成,让成片在一次导出中就能干净落地", which is that four kinds of audio track are generated with the picture rather than added afterwards. On length it says "单次生成即可获得完整 16 秒视频,表达更充分、叙事更连贯——减少拼接、避免节奏断裂", that one generation returns a complete sixteen-second video. On control it says "从运镜到节奏,你可以精确指挥每一个叙事节点;逐帧级控制帮助你落准关键时刻的时机、重音与情绪节拍".
The same page lists four highlights: synchronised audio and picture, output in English, Japanese and Chinese, a design aimed at comic drama, film and short drama, and natural multi-character dialogue.
The developer documentation adds a shorter description of the same model, quoted here in the original: viduq3 "支持智能切镜,支持音画同出,多机位的一致性更出色" — smart shot switching, audio generated together with the picture, and better multi-camera consistency.
Everything in this section is quoted from the publisher. Nothing on this page is a statement by Shengshu Technology, and this page has not been reviewed or endorsed by them.
Quotations are left in the publisher's original language, because translating a specification is one of the ways numbers drift.
What this site's service publishes for the same product
The workspace does not read the publisher's documentation. It reads this site's own service catalogue, which answers with one entry per model. The entry for this page is named Vidu Q3, sits in the family the catalogue calls Vidu Q3, and is marked available and callable. It declares one mode, reference to video, and the service's own one-line description of it reads "Vidu Q3 官方参考生视频模型,支持主体或 1~7 张参考图生成 3~16 秒视频,支持音画同出".
A floor is not a quote. The server totals the charge in points for the mode and settings you choose before a task starts, and that total is the one that counts.
| Field the service publishes | Value |
| Modes declared | reference to video, and nothing else |
| Task type behind the mode | reference-to-video |
| Published floor | 5 points per second |
| Billing mode | parameters plus duration |
| Input kinds | text and image |
| Reference-image limit | 1 to 7 |
| Duration | 3 to 16 seconds, default 5 |
| Resolutions | 540p, 720p, 1080p, default 720p |
| Aspect ratios | 16:9, 9:16, 1:1, default 16:9 |
One mode sold here, four modes in the publisher's line
This is the central fact about this page, and it is the reverse of the usual problem. The publisher documents four ways to make video with the Vidu Q3 line: from text, from an image, from a pair of keyframes, and from references. The service's entry for this page declares exactly one of them. The other three exist in the same catalogue, but as separate products under their own names, and none of them is reachable from this page.
So the workspace can express three modes this product does not sell, which is the opposite of a model whose capabilities are hidden by the interface. A visitor who opens Vidu Q3 from this page gets the reference-to-video form and nothing else, because that is what the service returns.
It is worth naming the consequence plainly. If you came here expecting to type a scene and get a clip, that is text to video, and it is a different product in the same family.
| Mode | Sold by this page's product | Where the catalogue does sell it |
| Reference to video | yes | Vidu Q3, 5 points per second |
| Text to video | no | Vidu Q3 Pro and Vidu Q3 Turbo |
| Image to video | no | Vidu Q3 Pro, Vidu Q3 Turbo, Vidu Q3 Pro Fast |
| Start and end frame to video | no | Vidu Q3 Pro and Vidu Q3 Turbo |
Every parameter the service publishes matches the publisher's
This does not always happen. On the video pages this site has built before this one, the two sides disagreed about duration ranges, about how many resolution tiers exist, or about whether a reference limit was real. For Vidu Q3 reference-to-video they agree, field for field.
The duration agreement is exact rather than close. The publisher's parameter table names the model by name and gives its row: "viduq3:默认5秒,可选:3 - 16". The service publishes a number field with a minimum of 3, a maximum of 16 and a default of 5.
There is also a third source, and for once it does not muddy the water. This site's source code carries a local family table whose entries are used only when the service publishes no value for a field. For this family it declares reference to video as the sole supported mode, which is what the service declares too. Nothing in it is overridden and nothing in it is inert.
The local table is mentioned because on earlier pages in this series it disagreed with both sides. Here it agrees with the service.
| Setting | Publisher's documentation | This service's published form |
| Duration | default 5 seconds, range 3 to 16 | default 5, minimum 3, maximum 16 |
| Resolution | default 720p, with 540p, 720p and 1080p | default 720p, with 540p, 720p and 1080p |
| Aspect ratio | default 16:9, with 16:9, 9:16 and 1:1 on the image-reference path | default 16:9, with 16:9, 9:16 and 1:1 |
| Reference images | 1 to 7 | minimum 1, maximum 7 |
The duration list differs between the publisher's own two surfaces
The agreement above is between the publisher's API documentation and this service. The publisher's consumer product page tells a third story, and it is the one most visitors will have seen, because it is the page with the generator on it.
The two publisher surfaces are describing different endpoints, so they are not in conflict with each other: a widget on a consumer page and a parameter table in an API reference are different documents about different call paths. What they do establish is that a number read off the consumer page is not the number the API accepts, and that a visitor who compares this page's duration field against the product page will see a difference that has a reason.
The third row has no such reason behind it. It is a third party's list, and it matches neither of the publisher's.
| Source | Duration options it offers |
| Publisher's API documentation, reference to video | 3 to 16, default 5 |
| Publisher's consumer page, image-to-video widget | 1 to 16 |
| A third-party review audited for this page | 4, 8 or 16 |
| This service's published form | 3 to 16, default 5 |
Reference images, and the references the form does not carry
Reference to video is the mode this product sells, so the reference limit is the setting that matters most. The publisher documents 1 to 7 images and the service publishes the same limit, with a minimum of 1. Reference images guide the shot without fixing its first frame, which is what separates this mode from image to video.
The publisher's own request example for this endpoint uses three images and a single sentence of prompt: "Santa Claus and the bear hug by the lakeside." Reading the three files against that sentence shows the intended division of labour, because the files are a lakeside scene at night, a red plush Santa, and a grey plush bear. One image carries the setting and two carry the subjects.
The endpoint also accepts video references, and here the two sides part. The publisher's documentation allows 1 or 2 reference clips, at most one eight-second clip or two five-second clips, in mp4, avi or mov, with a pixel floor of 128 by 128 and an aspect ratio no wider than 4 to 1. The service publishes an image field named image_urls and no video field at all, so that half of the endpoint is not reachable from this workspace.
One further publisher note is worth carrying because it prevents a wrong expectation. The voice identifier parameter, which chooses the voice used in generated speech, is documented as not being generated for q3 reference models, so it is not something to look for alongside the audio settings below.
Off-peak costs less, and the 48-hour window is part of the same sentence
The publisher documents an off-peak submission mode for this endpoint, and the documentation states four things about it in a single block. Quoted in full: "错峰模式消耗的积分更低,具体请查看产品定价" — it costs fewer credits. "错峰模式下提交的任务,会在48小时内生成,未能完成的任务会被自动取消,并返还该任务的积分" — a task submitted this way is generated within 48 hours, and a task that does not finish is cancelled automatically with its credits returned. Tasks can also be cancelled by hand. And "q3模型在audio参数为true时,支持错峰任务;q2、q1、2.0系列模型在audio为false时支持错峰任务" — q3 models support off-peak when audio is true, while the older series support it when audio is false.
Those four statements are one fact, not four. Reading only the first would give the impression of a straightforward discount, when what is actually being bought is a cheaper price in exchange for waiting up to two days. The direction of the audio coupling also reverses between generations, which is easy to read backwards.
This service's entry for this page does not publish a field for off-peak submission. The capability block belonging to a different product in the same family does carry a flag saying off-peak is supported, which is why the feature is described here at all, but the form behind this page's mode has no such control.
No wording on this page describes off-peak as simply cheaper, because the 48-hour window is published in the same paragraph as the discount.
Audio here is a capability the service declares and a control it does not publish
This is the most tangled of the disagreements, and it resolves into three separate facts that are all true at once.
First, the service declares audio. The capability block for this product says audio is supported, is generated, defaults to on, and has a toggle.
Second, the publisher also treats audio as on by default for this model, though its own parameter table states the opposite before correcting itself one line later. The parameter is described as "默认false" and then, immediately below, the same table says "viduq3、viduq3-turbo默认audio为true,其他模型为false". For this model the second line is the operative one. A second endpoint documented on the same page gives audio a default of true outright, and adds that outside subject-library calls only q3 models support the parameter at all.
Third, the service's published form for this mode contains no audio field. The fields are prompt, resolution, aspect ratio, duration and reference images. The workspace builds its controls from that form, and it renders an audio switch only when the form contains a field named generate_audio. It does not, so no audio switch appears.
The honest description is not that sound is off. It is that this workspace gives you no control over it, while the service's own capability block says a toggle exists.
| Publisher's audio parameters | What it does | Present in this service's form |
| audio | generate sound with the picture, or output a silent video | no |
| audio_type | choose all, speech only, or sound effects only | no |
| bgm | add background music, defaulting to off | no |
| voice_id | pick the speaking voice; not generated for q3 reference models | no |
What other pages for this term get wrong
Four pages were measured for this one, and the disagreements are worth stating because a reader comparing them will meet these numbers. They are listed as discrepancies rather than as errors, since a page can be describing a different surface than the one it appears to describe.
The reference-image claim is the one that would change a decision. Four images against seven is a difference a creator would notice on the first attempt, and the publisher's number is the one with a page behind it.
Something else about these pages is more informative than any single claim. Of the four audited, the two editorial reviews run to more than two thousand words each with dozens of images between them, and between them they embed no video at all. The only page in the set serving moving footage is a lookalike domain that is not the publisher. A page about a video model that shows no video is the gap this gallery exists to fill.
| Claim | Where it appears | What the publisher documents |
| At most 4 reference images | a third-party review of reference to video | 1 to 7 images |
| Duration options of 4, 8 or 16 seconds | a Chinese tool directory | 3 to 16 seconds, default 5 |
| Models named Q3 Drama and Q3 Ad | a third-party review | neither name appears in this site's catalogue |
| Per-second prices in dollars | a third-party review | the publisher prices this endpoint per second in its own units |
What this gallery shows, and what it deliberately leaves out
The publisher's developer documentation contains no generated output for this model at all. Across all seven archived pages there are no video elements and no clip files, and the only images present are seven files that appear as inputs inside request examples. The model's own output is on the consumer side, and it took two steps to reach, because the consumer site is assembled in the browser and a plain fetch of it returns a shell.
Reading the navigation out of the rendered consumer page produced a dedicated Vidu Q3 page, and that page serves four clips, every one of which downloaded. Three are published here as clips, because their length and their movement are the evidence: a ten-second two-person dialogue, a fifteen-second unbroken scene in a baseball stadium, and a five-second clip that cuts between three camera setups on one scene. The fourth is a hundred-and-sixty-second showreel, too long to publish, so five stills were sampled across it instead.
Two reels from the consumer home page were downloaded and then set aside. Their own file names are S11 and Q4-1, the page does not caption either as Vidu Q3, and the same site's navigation lists S as a separate line. Neither is attributed to this model.
Three of the stills published here are inputs rather than outputs: the lakeside scene and the two plush subjects from the publisher's own request example. They are labelled as inputs wherever they appear, and they are here because reference to video is the one mode this product sells.
Four further publisher images were retrieved and not published, and the reasons are recorded rather than left implicit. The image-to-video still, the start-end keyframe pair and the scene-template image all belong to sibling products selling modes this page's product does not, and the opening frame of the showreel is a near-black card carrying the publisher's own marketing sentence, which is copy rather than a generation.
Questions
Is this the official Vidu Q3 page?
No. It is a page on this site about the model. The publisher's own pages are named as sources throughout, and this page has not been reviewed or endorsed by Shengshu Technology.
Which modes can I actually run here?
Reference to video, and only that. Text to video, image to video and start-end to video are documented by the publisher for the same model line, and this catalogue sells them through separate products in the same family.
How many reference images can I attach?
Between 1 and 7. The publisher documents 1 to 7 and this service publishes the same minimum and maximum, which is worth checking because a third-party review for this term states a ceiling of four.
Can I attach a reference video as well?
Not from this workspace. The publisher allows 1 or 2 reference clips, at most one eight-second clip or two five-second clips, but the form this service publishes for the mode has an image field and no video field.
Why does the page show clips when the technical documentation has none?
Because the two surfaces carry different things. The developer documentation has no media output for this model at all, while the publisher's consumer Q3 page serves four clips, all of which downloaded and are used here. That page is assembled in the browser, so it has to be rendered rather than fetched.
Are the pictures of the lake, the Santa and the bear examples of what the model made?
No, and they are labelled as inputs for that reason. They are the three images in the publisher's own request example, published so the shape of a reference-to-video request is visible. The generated clips are the three items above them.
Does off-peak submission just cost less?
No. The publisher says an off-peak task costs fewer credits, and in the same paragraph says it is generated within 48 hours and is cancelled automatically with credits returned if it does not finish. For q3 models off-peak is also only available when audio is true, which is the opposite of the older models.
Can I switch the sound off here, or choose the voice?
There is no audio control on this page's form. The service's capability block declares audio as generated and defaulting to on with a toggle, but the published form carries no audio field, and the workspace renders its audio switch only when the form contains one. The publisher's voice-identifier parameter is documented as not being generated for q3 reference models in any case.
What does it cost?
The server quotes it in points for the mode and settings you chose, before the task starts. The catalogue publishes a floor of 5 points per second for this mode. That floor is not a quote and this page does not treat it as one.
The publisher's product page offers durations from 1 second, so why does the form here start at 3?
Because they are different endpoints. The widget on the consumer page is the image-to-video surface and lists 1 to 16. The API reference for reference to video documents 3 to 16 with a default of 5, and that is what this service publishes.