Grok Imagine Prompt Guide: Video 1.5 and Image 2.0 Prompts That Work, What Our Tests Found, With Examples You Can Run (2026)

Grok Imagine Video 1.5 is fast (under a minute a clip in our test) and makes its own soundtrack, so a good prompt describes the motion and the sound, and fences in the script: in our test it said the scripted line word for word and then added a reply nobody wrote. Grok Imagine Image 2.0 set every word right on two ads and made one of the most legible YouTube thumbnails in our tests, and was also the slowest and most expensive of nine image models on everyday jobs.

This guide covers video prompts, dialogue, starting from an image or references, image prompts, what to avoid, and examples you can run.

A note on who wrote this: GenMagic runs Grok Imagine Video 1.5, Video 1.5 Lite and Image 2.0. xAI publishes capability docs but no prompting guide, so the advice here comes from those docs, read November 10, 2026, and from our own tests, linked where used. Prices and plans: is Grok Imagine free? and Grok Imagine API.

Table of contents

1. Video prompts
2. Dialogue
3. Starting from an image or references
4. Image prompts
5. What to avoid
6. Examples you can run
7. Questions

Video prompts

• Lead with the motion: what moves, how, in what order. xAI's own example for an image-to-video clip is all motion: "Make the water crash down and slowly pan out the camera".
• Then the shot and the place: shot size, setting, light.
• Then the sound: "wind and surf", "market chatter", or "no music".
• Length: up to 15 seconds in one clip, by xAI's docs.
• One action per clip; a scene is several clips.

In our video model test, Grok Imagine Video 1.5 made a clean leap, a mid-air catch and a landing with sand thrown up, in under a minute, and returned its soundtrack at its price without sound.

Dialogue

• Quote the line and name the speaker, as with any video model.
• Fence the script in when it is approved: add "she says only this line; no one else speaks". In our test, Grok Imagine said the barista's line word for word and then gave the customer an unscripted reply, "Thanks. Looks perfect."
• Check the soundtrack before you use a clip for a client.

Starting from an image or references

• Opening image: the picture "becomes the starting point for the generated video"; describe only what moves.
• References: Grok Imagine Video 1.5 takes up to 14 reference images and 3 voice references, and can pin first, last and middle frames, by xAI's docs.
• Continue a scene: in the GenMagic studio, Continue starts the next clip on the last frame of a finished one.

Image prompts

• Exact text in quotes, with its place and look: huge bold yellow text with a thick black outline on the left reading exactly "I DROVE IT 1,000 MILES". Grok Imagine Image 2.0 set every word right in our ad test.
• Thumbnails: say "readable at small size" and keep the title short; in our thumbnail test it made one of the three most legible layouts.
• Edits: up to 5 source images in one request, for combining subjects or moving a style, by xAI's docs.
• Big images: it returned the largest images in our ad test, up to 1584 by 2816 pixels for a vertical ad.
• Describe only what a person has: in our headshot test, it added glasses after the prompt asked it to keep them.

On simple jobs, our image generator test found it the slowest and most expensive of nine models with no visible gain, so it earns its price on text-heavy thumbnails and ads rather than everyday pictures.

What to avoid

• Open-ended dialogue when the words matter: fence the script in, then listen.
• Several scenes in one prompt: split them into clips.
• Long text in a small space: keep titles short.
• Real people and brands without rights: not yours to use. A refused clip costs nothing on GenMagic.

Examples you can run

Each link opens the GenMagic studio with the model and the prompt ready; you see the price before you generate.

A fast scenic clip (Grok Imagine Video 1.5 Lite): "Wide shot, photoreal: waves crash against a black rock coast at sunset, spray rises in slow motion, a lone seabird glides through the frame, the camera slowly pans out to show the whole cove. Warm orange light, wind and surf sounds, 8 seconds." Open in the studio

A spoken line (Grok Imagine Video 1.5): "Medium close-up: a street musician in a red beanie plays a short, upbeat melody on a ukulele in a busy evening market, smiles at a passer-by and says "This one's for you." Warm string lights, handheld camera, market chatter, 6 seconds." Add "no one else speaks" if the line must stand alone. Open in the studio

A thumbnail with a title (Grok Imagine Image 2.0): "A YouTube thumbnail, 16:9: a bright red sports car parked on a mountain road at sunrise, huge bold yellow text with a thick black outline on the left reading exactly "I DROVE IT 1,000 MILES". High contrast, readable at small size." Open in the studio

Prices for every Grok Imagine model: Grok Imagine Video pay as you go and Grok Imagine Image pay as you go. Prompt guides for other models: Kling 3.0, Veo 3.1, Seedance, Nano Banana, GPT Image.

Teaching AI video or images on YouTube, in a course or in n8n templates? The creator program pays 15% in cash of what the people you bring pay in their first month.

Questions

Is there an official Grok Imagine prompt guide? Not as of November 10, 2026: xAI's docs describe what the models do (up to 15-second clips, image-to-video, up to 14 references and 3 voice references on Video 1.5) but give no prompting guide.

How do I stop Grok Imagine adding lines to my dialogue? Write the exact line in quotes, name the speaker, and add "no one else speaks"; then check the soundtrack. In our test it added an unscripted reply.

How long can a Grok Imagine video be? Up to 15 seconds in one clip, by xAI's docs; continue from the last frame for longer scenes.

Is Grok Imagine good at text in images? Grok Imagine Image 2.0 set every word right on two ads in our test and made one of the most legible thumbnails.

Is Grok Imagine fast? Video 1.5 finished in under a minute in our test, among the fastest of eight video models; Image 2.0 was the slowest of nine image models on simple jobs.