Prompt patterns across Grok Imagine Video models
For text-to-video, state the subject, action, camera, and light, then describe the ending that should still be visible when the clip stops. Keep one main action and one camera intention so motion failures are diagnosable. For an opening image, describe only the motion that should develop: what remains fixed, what moves, how the camera reacts, and which visual details must survive. Do not ask the prompt to redesign the source and preserve it at the same time. For reference work, assign a clear role to each reference instead of repeating visual detail. Name the images in order, identify which one controls the person, garment, object, location, or palette, and explain how those roles meet in a single shot. One or two precise references can be clearer than seven conflicting sources. Put negative constraints after the positive shot direction and reserve them for visible failure risks such as unwanted logos, duplicate subjects, extra limbs, accidental captions, or a camera cut. When comparing two prompts, change one variable and keep source images, duration, resolution, and framing stable.
Strengths and limits by model
Grok supports flexible text, opening-image, and multi-reference starting points. Text and image workflows can use the higher output tier shown by the model route, while reference mode is limited to its displayed 480p or 720p choices. The opening image owns framing in image-to-video, so a separate aspect-ratio promise should not be inferred. Reference mode accepts images, not a source video, audio track, or generic document. More references increase direction complexity and should not be treated as automatic consistency. Evaluate opening readability, subject identity, motion, anatomy and geometry, unwanted text or watermarks, and the final beat separately. The family page does not claim native audio, direct xAI execution, guaranteed lip sync, perfect typography, or controls absent from the selected mode. Queue behavior, availability, and exact pricing can change without changing the creative contract, so use the live model details and pre-render quote. For a high-stakes likeness or product, run a short diagnostic first, inspect every frame, and keep a human approval step before publication.
Pricing notes (varies by model)
Use the quote shown before generation; this copy owns no provider rate or finished total. Start with a short 480p or 720p reference diagnostic, or a 720p text or image test, before moving an approved direction to the higher settings offered by that mode. The reference quote must receive the actual number of images rather than a generic text-video assumption. Compare cost only when mode, duration, resolution, and reference count match. Keep rejected attempts in the production record so speed and stability are not judged only from successful examples.
