Prompt engineering for synthetic deep-sea mining video: A controlled evaluation of four text-to-video systems

Abstract Deep-sea mining video is scarce, and low illumination, sediment plumes, and moving equipment challenge visual monitoring. We evaluated four text-to-video systems on six VBench dimensions, first in an archived 300-video system-level comparison and then in a 240-video extension crossing four prompt conditions, five scenarios, and three requested seeds. Within-system prompt contrasts varied by metric; detailed or negatively constrained prompts did not consistently raise VBench scores. All 15 Wan pairs generated with and without a negative-prompt request had pixel-identical decoded frames, leaving that contrast uninformative for this service. In a separate 60-video review with system identities masked, three raters’ majority labels identified at least one domain violation in 37 clips (61.7%; descriptive Wilson 95% interval, 49.0–72.9%); agreement differed by criterion. These results support VBench scoring and domain-rule review as complementary first-stage screens, while physical fidelity and downstream utility remain to be tested.