Grok Imagine Prompting: The Video 1.5 Guide

TL;DR: Grok Imagine prompting splits cleanly by what you are making. The current video model is grok-imagine-video-1.5, it runs in the Grok app and through a real public API, and clips are 1 to 15 seconds with audio generated by default. Pick the right request mode first, then write the shot.

What is Grok Imagine, and which version are you actually on?

Grok Imagine is xAI's image and video product. The video model behind it, as published on xAI's own models page, is Grok Imagine Video 1.5, model id grok-imagine-video-1.5, with the aliases grok-imagine-video-1.5-preview and grok-imagine-video-1.5-2026-05-30.

The version numbers are worth two minutes of your attention, because they contradict each other in a way that trips people up.

• xAI's news index lists Grok Imagine 1.5 Preview on 3 June 2026, Grok Imagine Video 1.5 generally available on 16 June 2026, and Imagine Video 1.5 with References on 31 July 2026.
• The consumer app tells a different story. On 28 August 2026 the served client configuration on grok.com carried the upgrade line "Imagine 1.0 is here" and a benefit reading "Higher usage on Grok Imagine 1.0".

So the product is branded 1.0 and the video model is 1.5. When someone says "Grok Imagine 1.5" they mean the model. That is the number that matters for prompting, because it is the number attached to every capability xAI documents.

Is Grok Imagine video app-only, or is there an API?

There is an API, and this is the single most common error in articles on this keyword. Video generation runs at POST /v1/videos/generations on api.x.ai, with POST /v1/videos/edits and POST /v1/videos/extensions alongside it. xAI shipped the Imagine API on 28 January 2026 and has been adding to it since.

It is asynchronous. You submit a request, get back a request_id, and poll GET /v1/videos/{request_id} until status reads done. The Python and Vercel AI SDKs wrap that loop for you; the REST path does not.

That matters for prompting more than it sounds. The app gives you a picker. The API gives you duration, aspect_ratio, resolution, generate_audio, reference_images and reference_audios as explicit parameters, which means anything you can express as a parameter should leave your prompt entirely. Writing "in 16:9" into your prompt text on a surface that has an aspect_ratio field is wasted tokens and an invitation for the model to render letterboxing into the frame.

One more app-only detail: on 28 August 2026, grok.com's configuration exposed three video prompt modes, Custom, Normal and Spicy. The word "spicy" appears zero times across the entire docs.x.ai corpus. It is an app control, not an API concept, and no prompt syntax will summon it through the API.

The API also tells you why a generation failed, which the app does not. A failed job returns an error.code, and the one that concerns prompt writers is invalid_argument, which xAI's own table describes as covering "an unsupported duration, an invalid image or video input, a prompt that is too long, conflicting request modes, or content blocked by moderation". Note that four of those five are request-shape problems rather than wording problems. Separately, every completed video carries a respect_moderation boolean, so a generation can succeed technically and still come back filtered. Check the flag rather than assuming a returned URL means a usable clip.

Which request mode does your shot actually need?

xAI's video endpoint decides its mode from which fields you set, and only one mode can be active per request. Getting this wrong is the most common reason a perfectly good prompt produces a disappointing clip.

Two details from those pages change how you write. First, on grok-imagine-video-1.5, "text-to-video uses text-to-image then image-to-video under the hood: the model generates a first frame from your prompt, then animates it." Your text-to-video prompt is therefore doing two jobs at once, and the first half of it is really an image prompt. Second, reference images "incorporate specific people, objects, clothing, or other visual elements without locking the first frame", which the same page explicitly contrasts with image-to-video. If you need a specific opening frame, that is image-to-video, not references.

Here is the skeleton I use across all three modes. It exists to make sure the frame is decided by you rather than by the model.

[Subject and what it is doing]
[Setting, time of day, and one concrete detail that anchors the place]
[Camera: shot size, lens feel, and one movement]
[Light: source, direction, quality]
[Motion pacing: how fast anything moves]
[Audio: what should be heard, if anything]
[Grade or film reference, one clause only]

How do you prompt a shot that does not exist yet?

Text-to-video is for shots you have no plate for. Because the model builds a first frame from your words before animating it, front-load the composition and put the motion after it. Vague nouns produce a generic first frame, and no amount of camera language rescues a generic frame.

A ceramic espresso cup on a scratched steel counter in a closing cafe, one overhead pendant lamp still on, everything else dark. Slow push in from a wide two-thirds shot to a tight three-quarter. Steam rises and curls left. Warm tungsten key from above, deep shadow behind. Nothing else in the frame moves. Faint hum of a refrigerator. Muted, slightly desaturated grade.

A single red kite tangled in the top branches of a bare oak, seen from below against a flat grey sky. Static wide shot, no camera movement. Wind gusts twice: the kite snaps taut, then slackens. Overcast, no visible sun, soft shadowless light. Wind noise, no music. Cold, high-contrast grade.

An open notebook on a train seat beside a rain-streaked window, handwriting visible but not readable. Locked-off medium close-up. The pages lift and settle as the train moves; light from the window pulses as poles pass. Grey daylight, blue-shifted. Rhythmic rail noise underneath. Naturalistic grade, slight grain.

A hand-thrown clay bowl on a potter's wheel, mid-spin, wet clay glistening. Overhead top-down shot descending three inches. Wheel keeps a constant speed; two fingers enter frame from the right and shape the rim. Single soft window light from the left. Wet clay sound and a low wheel motor. Warm, natural grade.

Note what is missing from all four: no resolution, no aspect ratio, no duration, no "4K", no "cinematic masterpiece". Those either belong in parameters or belong nowhere. The camera clause is the one place worth spending words, because it is the only instruction in the prompt that the model cannot infer from the subject.

How do you prompt a still photo into motion?

Image-to-video locks your first frame. The image has already decided the subject, the composition and the light, so your prompt should stop describing them and start describing change. The most common failure here is re-describing the photo, which gives the model permission to reinterpret it.

Also worth knowing: the output defaults to the input image's aspect ratio, and passing aspect_ratio overrides that and stretches the image to fit.

Hold the composition exactly. The woman turns her head six inches to the left and her eyes follow. Hair moves slightly. Camera stays locked. Nothing else changes.

Keep the frame. Push in slowly over the full clip, ending about fifteen percent tighter. The candle flame flickers twice. Shadows on the wall shift with it. No other movement.

Hold the plate. The steam above the bowl thickens and drifts right. A hand enters from the bottom of frame, sets down a spoon, and withdraws. Camera static throughout.

Keep the subject and lighting. Slow parallax dolly to the right, about a foot, revealing more of the window behind him. He does not move. Dust drifts through the light beam.

How do you keep the same face, product and place across clips?

Reference-to-video is the consistency tool, and it is the mode most people never find. xAI's 31 July 2026 announcement puts it as plainly as anyone could: "Each reference image locks one thing in place — a face, a product, a location." The same page adds, "Up to seven references per generation."

You address the references inside the prompt by index tag. This is the closest thing Grok Imagine has to a syntax, and it is the part worth memorising.

The person from <IMAGE_1> stands in the kitchen from <IMAGE_2>, holding the bottle from <IMAGE_3> at chest height, label toward camera. Medium shot, slight push in. Morning window light from the left. They look down at the bottle, then up at the camera. Room tone only.

Keep the character from <IMAGE_1> exactly. New setting: a rain-soaked car park at night, sodium lights overhead. Wide shot, camera tracks with them left to right as they walk. Wet asphalt reflections. Footsteps and rain, no music.

Same location as <IMAGE_1>, same camera position, but the person from <IMAGE_2> now sits at the table instead. Static medium shot. They pour from the jug, set it down, and look off frame right. Available light only.

The model from <IMAGE_1> wears the jacket from <IMAGE_2> over plain black trousers. Three-quarter full-length shot on a seamless grey backdrop. They turn ninety degrees to their left, pause, and turn back. Soft studio key with a large source camera-left. Fabric should move with the turn.

The product from <IMAGE_1> sits on the surface from <IMAGE_2>. Slow ninety-degree orbit around it, ending on the label. Single hard key from the upper right, one bounce card filling the shadow side. No hands, no people. Faint room tone.

If character consistency is the whole problem you are solving, the cross-model treatment in keeping a character consistent across clips covers what transfers between vendors and what does not.

How do you prompt for sound and dialogue?

Audio is generated in the same pass, and it is on by default. xAI's June announcement describes it directly: "Sound effects, ambience, and dialogue are generated in the same pass and land on the action." Pass generate_audio as false if you want silence, because a stray generated sound bed is much harder to remove later than to never request.

For voice, reference-to-video on grok-imagine-video-1.5 accepts up to three preset voices by voice_id, drawn from the same roster as xAI's text-to-speech API, and you address them as <AUDIO_0>, <AUDIO_1> and <AUDIO_2>. Caller-supplied voice clips are limited to trusted partners on request, so a preset voice is what most people get.

Close-up on the kettle as it comes to the boil. Rising whistle, then a click as it switches off, then room silence. No music, no ambience beyond a quiet kitchen.

The person from <IMAGE_1> speaks with the voice from <AUDIO_0>, direct to camera: "We shipped it on Friday and nobody noticed. That was the point." Medium close-up, static. Soft key from camera left. Room tone under the dialogue, nothing else.

Two people at a cafe table. The one on the left speaks with the voice from <AUDIO_0>, the one on the right replies with the voice from <AUDIO_1>. Overlapping cafe ambience: cups, a grinder, low conversation. Two-shot, slight handheld drift.

A workshop bench, no people. Only sound design: a plane shaving wood in three long strokes, each landing on the visible motion, then a shaving dropping to the floor. No music. No voice.

The pattern that works across all four: name the sound, then name the visible action it should land on. Sound described in isolation tends to arrive as a generic bed underneath the clip rather than as an event inside it.

How do you change one thing in a clip you already like?

Video editing takes an existing video and a prompt, and xAI's page emphasises "strong scene preservation, modifying only what you ask for". Write these as instructions, not as descriptions. One change per request, then chain.

The limits are firm and worth reading before you plan a workflow. The input must be .mp4. The maximum input length is 8.7 seconds. Duration, aspect ratio and resolution are not accepted; the output inherits the input's, capped at 720p.

Give the woman a silver necklace. Change nothing else.

Change the colour of the jacket to deep burgundy. Keep the fabric texture, the folds and the lighting exactly as they are.

Replace the sky with heavy overcast cloud. Match the existing light direction on the subject. Do not alter the foreground.

Remove the coffee cup from the table. Fill the space with the same wood surface. Leave everything else untouched.

How do you get a scene longer than 15 seconds?

You extend. POST /v1/videos/extensions continues a clip from its last frame, and the docs are specific: the input must run between 2 and 15 seconds, the extension range is 2 to 10 seconds with a default of 6, and the duration parameter "controls the length of the extended portion only", not the total output. A 10-second input plus a duration of 5 returns 15 seconds.

Write the extension prompt as what happens next, in continuity terms. Restating the scene tends to produce a visible seam.

Continue the same shot. The camera keeps drifting right at the same speed. She finishes the sentence, looks down, and picks up the pen. Same light, same room tone.

Continue from the last frame. The car passes out of frame to the left, and the camera holds on the empty road for two seconds before a second car enters from the right. Same grade, same time of day.

Continue. The push in stops. Hold the tight shot for the remaining time as the steam thins out and the light dims slightly. No new elements enter the frame.

For an approach to planning long generations shot by shot rather than fighting a duration cap, structuring a 20-second generation works through the same problem on a model built for it.

What does xAI actually publish about duration, resolution and aspect ratio?

Everything below is from docs.x.ai, accessed 28 August 2026. Nothing here is inferred.

That headline rate needs a footnote. xAI's models page and pricing page both list grok-imagine-video-1.5 at a flat $0.080 per second. The model's own detail page lists per-second output pricing that scales with resolution: $0.08 at 480p, $0.14 at 720p, $0.25 at 1080p. Both were live on 28 August 2026 and neither references the other. Budget from the higher figure.

The other cost lever nobody plans for: 480p is the default resolution. If you never pass resolution, you are testing at the cheapest setting, which is exactly what you want while you are iterating on wording.

Do not generate real, identifiable people

This one is short and non-negotiable. Do not prompt Grok Imagine for real, identifiable people.

xAI's Acceptable Use Policy, effective 14 August 2026, prohibits "Deceptively impersonating a real person" and "Depicting a person in a defamatory manner or portraying them in a false light", among other restrictions on likeness and privacy. The same document prohibits "Stripping, altering or circumventing embedded provenance metadata or watermarks", and xAI's FAQ states plainly that "There is no setting to remove the watermark."

Two things follow. First, the platform's rules apply whether or not a given prompt happens to succeed; a generation that is not blocked is not a generation that is permitted. Second, and independently, publicity, likeness and defamation law applies to you as the person publishing the output, in whatever jurisdiction you publish it. Read xAI's Acceptable Use Policy yourself rather than trusting anyone's summary of it, including this one.

There is a related trap in the app. xAI's FAQ notes that "Enabling NSFW does not turn off moderation", and that the company deliberately does not publish its moderation rules. So you cannot reason your way to what is allowed from observed behaviour. The policy is the document; the model's output is not evidence of anything.

Does xAI publish a prompting guide for Grok Imagine?

No, and saying so is more useful than pretending otherwise. As of 28 August 2026 the only pages titled "Prompting Guide" on docs.x.ai cover the speech-to-speech API and multi-agent research. There is nothing equivalent for images or video.

The absences matter as much as the presences. Searching the whole published docs corpus on 28 August 2026 returns zero occurrences of negative_prompt, and the video capability and REST reference pages contain no seed field and no style parameter. So three habits people carry in from other models do not apply here. You cannot suppress an element with a negative prompt, because there is nowhere to put one. You cannot reproduce an exact result by pinning a seed, because there is no seed to pin. And you cannot select a look by naming a preset, because there is no preset list. Everything you want has to be stated positively, in the prompt, as something present in the frame. If you are used to writing exclusions, our negative prompt support matrix shows which vendors actually expose the parameter and which only appear to.

The closest thing to official prompt advice is a single sentence in the 3 June 2026 preview announcement: "Direct the shot with natural-language prompts. Describe the camera move, the pacing, and the sound design, then set your resolution and clip length." That is a real instruction from xAI and it is the shape the skeleton above follows. Everything else you read about "Grok Imagine prompt syntax" is somebody's craft observation, ours included, and should be labelled as such.

Practically, that means the loop is: write one paragraph of plain prose at 480p, look at what the model actually gave you, change one clause, regenerate. Not a magic word list.

Where does Prompt Architects fit?

We do not generate video. Grok Imagine does that, and so do Veo, Kling, Seedance and Runway. What we generate is the prompt — a structured, reusable one, from a rough idea, which you then paste into whichever model you are using this week.

That is worth something specifically because Grok Imagine has no prompting guide. When there is no vendor spec to follow, the value moves to having your own house structure and reusing it. Video prompt generation is on our Advanced and Team plans, not Pro. The Free plan includes 5 prompt enhancements per day, forever, according to our FAQ page.

If you want the ready-made versions rather than a method, our Grok Imagine prompt template library has copy-paste sets by genre, and what xAI actually documents is the reference companion to this page.