How to prompt Grok Imagine Video 1.5: style labels, comic timing and sound cues, with a reusable template, three example prompts and the mistakes to avoid.
Grok Imagine Video 1.5, made by xAI, takes a prompt, a first frame or a set of reference images, and it always adds native audio. Because picture and sound come from one prompt, a strong prompt sets the tone up front, then backs it up with concrete cues for the eye and the ear. This guide covers how the model reads a prompt, a template, three example prompts and the mistakes worth avoiding.
A prompt alone gives you text-to-video. Add a first frame to fix the opening image. There is no last frame, so the prompt has to carry where the shot goes. With a first frame, the output follows the frame and the aspect ratio setting is ignored. Grok Imagine Video 1.5 References takes up to 7 reference images in place of a frame, for characters, props or places that need to match other material.
Clips run from 1 to 15 seconds. The main lane renders 480p, 720p or 1080p, with ratios from 16:9 to 9:16, and the References lane stops at 720p. Native audio is always on, so every prompt is also a sound brief.
[Style or genre label] [shot size]: [subject] [action], [reaction or secondary beat], [camera move]. [Texture or look cues]. Sound of [effect] and [ambience].
Put the label first: "Retro 1980s VHS look:", "Comedic wide shot:" or "Documentary style close-up:". Then describe one action and the camera, and finish with sound. Since you cannot set a last frame, end the action on the moment you want the shot to land on.
Made in Cinemagiq. The exact prompt is under each one.
PromptRetro 1980s VHS look: a car drives down a palm-lined highway at sunset toward a giant synthwave sun, scan lines and slight tape wobble. Sound of the engine and wind.
PromptComedic wide shot: a knight in full armor tries to sit on a tiny cafe chair, the chair creaks and the barista stares. Sound of creaking wood and cafe chatter.
PromptDocumentary style close-up: a red panda climbs along a mossy branch in light rain, the camera follows smoothly. Sound of rain on leaves.
Example 1 leads with a period look, "Retro 1980s VHS look:", then backs it up with concrete texture cues at the end: "scan lines and slight tape wobble". The label names the era, and the cues tell the model what that era looks like on screen. "Sound of the engine and wind" covers the car and the open road.
Example 2 sets the tone and the framing at once with "Comedic wide shot:". A wide shot leaves room for the gag and the reaction in one frame. The beats run in order ("tries to sit on a tiny cafe chair, the chair creaks and the barista stares"), and the sound ("creaking wood and cafe chatter") underlines the joke.
Example 3 borrows a genre, "Documentary style close-up:", and pairs it with a camera note that fits that genre: "the camera follows smoothly". The subject is a natural one, a red panda on "a mossy branch in light rain", and the sound stays with the weather: "Sound of rain on leaves".
For more shot types, moves and lighting terms, see the prompt engineering cheat sheet.
Yes. Native audio is always on, so every clip has sound. Describe it in the prompt, for example 'Sound of rain on leaves', so the audio is something you chose for the shot instead of something left entirely to the model.
It takes a first frame, but not a last frame. With a first frame, the output follows the frame and the aspect ratio setting is ignored, so compose the still at the ratio you want. Grok Imagine Video 1.5 References takes up to 7 images instead of a frame.
From 1 to 15 seconds, at 480p, 720p or 1080p. Very short lengths suit inserts and quick reaction beats, and longer ones leave room for an action to play out. Grok Imagine Video 1.5 References renders at 480p or 720p.
Open with a style or genre label, such as 'Retro 1980s VHS look:' or 'Documentary style close-up:', then back it up with concrete cues like scan lines or a smooth follow. One clear look per prompt is easier to control than several mixed ones.
Run Grok Imagine Video 1.5 and the other leading models in one AI-native workspace.