Reference to Video

Reference to video uses one or more images as a guide rather than as the literal first frame. The reference tells the model what a subject, a character or a visual style should look like, and the prompt describes the clip to build around it. The output does not have to begin with your picture; it has to look like it belongs with it.

That distinction matters when you need consistency across several clips. A recurring character, a product with a specific shape or a house style is far easier to hold across shots by feeding the same reference each time than by re-describing it in words. Clean references work best, with the subject clearly visible and the background quiet, because a busy photograph gives the model too many competing cues about what it should carry over.

On E2X this category is served by Google's Veo 3.1 Fast and Omni Flash. Both use the same call shape as everything else in the catalog: post a model slug with your input object to the jobs endpoint and poll the job id it returns. Requests are billed individually against a prepaid balance, so testing the same reference against both models is two requests and nothing more.