How to prompt Grok Imagine Video: the model that cannot be silent

Grok Imagine Video is xAI's video line, and the thing to know before you write anything is that its audio has no off switch. Picture and sound co-generate in one pass, and none of the models expose an audio parameter—so if your prompt does not say what the clip sounds like, the model decides for you. On LUVI it runs as text-to-video, image-to-video, reference-to-video, video editing and video extension, plus a v1.5 image-to-video model. On September 22, 2026 an eight-second clip quoted 800 credits at 480p and 1,120 at 720p.

How do I write the sound?

As its own clause, every time. LUVI's prompting guide for the family is blunt about the failure mode: audio is always on with no silence flag, so a prompt that does not specify sound gets silent or random audio. Write it deliberately instead—music, effects timed to the motion, room tone, or a spoken line.

For dialogue, keep the line short and put a delivery cue in front of it. Audio is the weakest layer in this model, and long speeches turn to mush. A sound-design clause at the tail is the tidiest place for everything else.

A lighthouse keeper shoulders open a heavy iron door at dawn, wind dragging hard at his coat. Locked camera, cold blue light across wet stone, grainy 16mm look, sharp focus throughout. A weary mutter: "Still standing." AUDIO: howling wind, the hinge groaning, distant surf.

What order should the prompt follow?

Timeline order, with the important thing first: subject, then action, then camera, then environment and lighting, then style, then audio. Grok's engine is autoregressive, which means earlier words carry more weight than later ones—so the action you would reject the clip for belongs near the front, not buried after three clauses of mood.

Thirty to sixty words is the working band. One primary action and one named camera move per clip: conflicting moves, a zoom stacked on a pan, break the motion apart.

Why did "no blur" give me blur?

Because negation is ignored on this line. "No X" is still tokens describing X, and the model has no negative prompt field to route them into. Every exclusion has to be rewritten as the presence of what you want: "sharp focus throughout" instead of "no blur", "an empty street at dawn" instead of "no people", "smooth clear skin" instead of "no blemishes".

The same discipline applies to motion. The model cannot infer magnitude, so "passing" and "passing quickly" are different instructions—use strong verbs and say how much: "sprints", "surges", "pitches forward".

Which shots does it handle badly?

Three, and knowing them saves credits. Large body dynamics come back jerky. Extreme close-ups and macro hand work distort. Long photorealistic holds go waxy as the seconds run on. Favor short beats, implied motion and reaction shots over sustained full-body action, and pick one aesthetic per clip—mixing anime and photoreal in one prompt averages them into neither.

What do the parameters do, and what do they cost?

Three, and two of them move the price.

• duration runs from 1 to 15 seconds, 8 by default. xAI's documentation gives the same range, and notes that video editing does not take a custom duration at all—an edited clip keeps the original's length, capped at 8.7 seconds. On LUVI that shows up as a flat price for the editing model rather than a per-second one.
• resolution is 480p or 720p. xAI's own docs mention native 1080p for v1.5 text-to-video, but on LUVI the Grok video models top out at 720p.
• aspect_ratio offers seven values, including 9:16 for vertical.

Length and resolution both scale the bill: the same eight seconds went from 800 credits at 480p to 1,120 at 720p. xAI describes the same shape—"video generation uses per-second pricing where both duration and resolution affect the total cost".

What changes with an image or references?

The frame stops being your job. With a starting image, composition, lighting, style and identity are locked, so write motion, camera and audio only, and close with a preservation clause when the face matters: "keep the face and outfit unchanged." With references, each one should carry a single distinct element—character, prop, background or style—and the reference-to-video model on LUVI takes up to seven images, with durations up to ten seconds rather than fifteen. For an extend, write only what happens next; recapping the source clip gives the model nothing.

How we checked

• Models: Grok Imagine Text-to-Video, Reference-to-Video and Video v1.5 Image-to-Video, with the editing and extension models read from the catalog.
• Settings: eight-second clips at 480p and 720p, 16:9.
• Date and what was read: September 22, 2026—the models' parameter contracts and live credit quotes in LUVI, LUVI's own prompting guide for the Grok video family, and xAI's video generation documentation.
• Results: 800 credits at 480p and 1,120 at 720p for eight seconds. No model in the family exposed an audio parameter, and none offered a resolution above 720p. No outputs were generated for this post, so it makes no claim about output quality.
• Drawbacks: the failure modes listed above come from LUVI's prompting guide rather than a published benchmark, and the 30-to-60-word band is a working range, not a measured optimum.

In LUVI

Pick a model from the Grok Imagine family page, write the audio clause before you generate, and let the feather button check the prompt against LUVI's prompting guide—it is the step that catches a stray "no" and an unnamed soundtrack. If you are used to models where audio is a toggle, the help center's list of which models generate sound with the video is worth a look first, because the habits differ family by family.

New accounts start with 200 credits, and verifying a phone and a card adds up to 1,500 more. You can create a LUVI account and hear what the model does with a properly written audio line.

Sources

• Video generation — xAI, official developer documentation. Accessed September 22, 2026.
• Imagine API overview — xAI, official developer documentation. Accessed September 22, 2026.