Most discussions of generated video stall on the model. The people who ship clips worry about something narrower: the moment a take leaves the generator and enters an edit, where it has to cut against real footage and survive a review. Vidu Q4 aims at that slot instead of at a whole scene. According to the product page, it returns clips of three to sixteen seconds at resolutions from 540p to 4K, working either from one starting image or from a required prompt plus one to twelve reference pictures.

That short ceiling is the useful part of the constraint. A few seconds is not a scene; it is b-roll, a transition, an insert, a plate. Judging such a clip by scene standards burns renders, and judging it by b-roll standards changes what belongs in the prompt. The five checkpoints below are the ones that keep a take from being rejected after somebody already signed off on it.

1. Reserve the delivery slot before writing a word

Frame shape and length are delivery decisions. They cost nothing to decide and a lot to discover late, because a clip that is beautiful in the wrong shape is simply the wrong clip.

Frame shape fixed first: 16:9 for a wide plate, 9:16 for a vertical cut, 1:1 or 3:4 for a feed placement.

Target length rounded to a length the tool supports, not the length in the brief.

Resolution chosen for the largest screen the clip will ever reach, and written next to the shot name.

2. Match the mode to the material you actually hold

The two entry paths ask for different inputs, and picking the wrong one yields a clip that is technically clean and editorially useless. Image to video starts from one still, and the still settles subject and composition before anything moves. Reference to video starts from a written prompt and one to twelve pictures of the same scene, which is the right entry when you have coverage but no single frame that represents the shot.

If one frame already states the composition, animate that frame.

If the subject only exists across several angles, supply the prompt plus the picture set.

Never use the reference path to rescue an unusable still, and never use the image path to imply motion the still cannot support.

3. Switch prompt expansion off when the wording is contractual

According to the product page, prompt expansion is available and enabled by default in the image mode, and is not offered in the reference mode. Expansion is helpful during exploration, when a literal reading of your sentence is probably worse than an interpreted one. It is a defect when a client wrote the sentence and expects the clip to match it.

Keep a note of which setting produced which take.

Read the expanded result as a separate take, not as the same take with better luck.

When wording is fixed, run with expansion off and judge motion alone.

4. Lock the seed the moment a take is approved

Approval without a record is not approval; it is a memory. The product page lists a seed field that accepts a blank value for a random result or a nonzero integer to reproduce a specific one, which means reproducibility is a choice you make at render time, not a favour you ask for afterwards.

Record the integer beside the take name, not in a chat thread.

Record mode, resolution and duration with it, since a seed alone does not describe the take.

Archive the approved file under those parameters rather than under a version label.

5. Sign off audio before anyone signs off the picture

Sound is the element reviewers notice last and reject first. The product page describes a toggle for generated audio, and in the reference mode optional audio references that it explicitly presents as references rather than voice cloning. That distinction is worth carrying into a client note, because the two are easy to merge in a review thread and hard to separate afterwards.

Decide whether the clip needs sound at all before rendering, not after.

Confirm what an audio reference is being used for, in plain words.

Check the clip muted as well, since most first views happen without sound.

The two-minute pre-flight

Run this before calling anything final, and run it in this order:

The slot is reserved: shape, length, resolution.

The mode matches the material on hand.

Expansion is set deliberately, not by default.

Seed and settings are written down.

Audio is decided, described and checked muted.

Nothing on that list is about prompt craft, and that is the point. Prompt quality decides whether a clip is interesting; these checks decide whether it can be used. A generator only needs to be predictable for the workflow to hold together, and predictability is mostly a matter of the settings being visible, repeatable and named the same way in the tool and in the review. Vidu Q4 is worth evaluating on that basis rather than on a demo reel: mode, resolution, duration and seed are the four fields a reviewer will ask about, and a clip that carries all four is far easier to defend than a clip that arrives with only a filename.