Grok Imagine Video 1.5: Seven References, Same Face and Voice

👁 116 reads

Imagine Video 1.5 gives you up to seven reference images per generation, each locking one element: face, product, or location. Add voice reference, and your character keeps the same face and voice across every scene. That’s the difference between a single clip and a series that holds together.

What’s New in This Release

How to Think About References - This Is Key

The easiest way to understand this is as three sliders, each of which you can lock independently:

• Keep the character, change the scene - same person in five different locations
• Keep the scene, change the character - different people in the same interior
• Keep both, change the action - same person in the same place doing something different

That third option is most valuable for a series. It lets you build a scene shot by shot, instead of generating five independent clips and hoping they match.

Ready-to-Use Prompts for Testing

1. Tripod Shot - Basic Test

Static tripod shot, no camera movement. A leather suitcase stands alone on an empty bus station bench at night. Steam drifts slowly across the frame from the left. Cold blue light from a single overhead lamp. 6 seconds, 16:9, 1080p.

Why tripod: camera movement is the fastest route to cheap-looking generation. The model then has to maintain geometric consistency of the scene in motion, and that’s where it most often fails. A static camera with movement in frame looks much better with the same model.

2. Character Across Multiple Shots

First generate an image of the character, then use it as reference:

Scene 1: Slow push-in on the character standing at a rain-streaked window, looking out. She does not move. 5 seconds.
Scene 2: Same character, now seated at a kitchen table, turning a coffee cup slowly in both hands. Static shot. 5 seconds.
Scene 3: Same character walking away down a narrow corridor, seen from behind. Static shot. 5 seconds.

In each generation, you provide the same character reference image. The face stays consistent.

3. Product in Motion

Static macro shot. A matte black cosmetic bottle on a white surface. A single drop of water runs down the label. Soft studio light from the left, one long shadow. No hands, no text, no people. 4 seconds, 1:1.

Four Rules We Learned from Our Own Clips

These are our observations from working with video generators, not manufacturer recommendations:

1. Write prompts in English, always. Non-English prompts give noticeably weaker results in every generator we’ve worked with.
2. Never ask for captions or text in frame. Video generators turn letters into gibberish. Add text in editing.
3. Four to six seconds per shot. Longer generations fall apart at the end - the character starts changing, the background drifts.
4. Tripod instead of camera movement. We’re repeating this deliberately because it’s the one change that gives the biggest improvement.

Via API

The model is called grok-imagine-video-1.5. The call accepts a prompt, list of reference image URLs, duration, aspect ratio, and resolution:

client.video.generate(prompt="...", model="grok-imagine-video-1.5", reference_image_urls=[...], duration=6, aspect_ratio="16:9", resolution="720p")

Image references, text-to-video generation, and native 1080p are in the API. Voice reference is available on request, meaning you need to ask for it separately.

What Not to Expect

• You won’t make a film with one click. This is a tool for shots, not scenes. Editing is still on your side.
• Audio requires checking. The manufacturer mentions better audio, but for serious work you’ll add your own anyway.
• Voice reference is not default in the API - available on request.
• Long shots still fall apart. That’s a limitation of the entire category, not this model.

Real-World Workflow

1. Generate the character as an image in an image generator and save it - that’s your reference for the entire series.
2. Do the same with the location, if it will repeat.
3. Generate shots of 4-6 seconds, each time with the same references.
4. Edit and add audio outside the generator.
5. Captions only in editing, never in the prompt.

This sequence seems obvious, but most people start by generating shots and only discover on the third one that nothing matches.

Common Questions

How many references can I provide?

Up to seven per generation. Each locks one element - face, product, or location.

Can you make video without a starting image?

Yes, text-to-video generation is generally available on grok.com/imagine and in iOS and Android apps.

What resolution?

Native 1080p for both text-to-video and image-to-video generation.

Will the character keep the same voice?

Yes, when you provide a character image and voice reference, both stay consistent across every scene. In the API, this feature is available on request.

How long can shots be?

You set the duration parameter in the call. From practice: four to six seconds gives the most stable result, longer generations fall apart at the end.