Contra Labs ran xAI's Grok Imagine through every phase of ad video production: ideation, mockup, refinement. It produced the most dramatic phase-over-phase improvement of any video model in the study.
Surge at refinement
By refinement, every major theme flipped positive. Motion Quality (+16), Usability (+23), Realism (+20), Prompt Adherence (+11). Win rate: 56%, ahead of every other video model we tested. Evaluators kept reaching for the same four words: organic, smooth, natural, controlled.
The most usable and quality generation, free of visual oddities and unrealistic elements.
The camera movement showcases the product well. The movements of the clouds in the sky and the breeze are well-matched.
The neon edge illumination is beautifully lit, the camera movement pans in slowly. The close-up shot is golden. A nice ad.
Where it earns the win
Refinement is the phase where creatives stop generating and start polishing. Smooth camera moves, natural product framing, controlled lighting transitions. Grok handles all three with consistency. Motion quality (30 vs 18 across competitors), usability (24 vs 18), and lighting (17 vs ~3) all sit clearly in the green.
Where the next gains pay off
Two signals from the data point to where additional work compounds.
Ideation is rougher. Net Realism (−15) and Scene Coherence (−8) take hits from technical errors: object duplication, hands appearing out of frame, physics that doesn't quite hold. Grok seems to need material to iterate on before it shines.
There are two hands appearing in the shot instead of one, and they are interacting with themselves in a very unnatural way.
The portafilter is duplicated upon lock-in. Coffee pours in the device's sink.
Refinement variance is high. 42% first-place finishes alongside 25% fourth-place. When Grok hits, it leads the field. When it misses, it misses by a lot. Smoothing that variance is where the gains compound.
The practical implication
Bring Grok Imagine in once you have a draft worth iterating on. It takes you from "almost there" to "client-ready" better than anything else we tested.
Knowing which model to use at each phase is now a real creative skill.