T2VAttackBench: A High-Quality Dataset for Adversarial Attack on Text-to-video Diffusion Models

T2VAttackBench is a textual benchmark intended for rigorous evaluation of adversarial robustness on text-to-video (T2V) diffusion models. While prior resources such as VBench provide a comprehensive benchmark, the limited video generation quality makes them suboptimal for robustness studies. The dataset focuses on two distinct evaluation objectives: semantic fidelity (how well generated video matches the prompt) and temporal dynamics (quality of motion and dynamics). A two-stage filtering process (model-centric selection and consensus filtering) ensures that all prompts yield semantically faithful and temporally coherent videos, thereby isolating degradation caused by adversarial factors rather than intrinsic model weaknesses.The final dataset includes 105 semantic prompts, describing visually rich and contextually coherent scenarios, and 52 temporal prompts, specifying fine-grained motion and dynamic video content. These curated prompts provide a reliable foundation for evaluating semantic fidelity and temporal consistency in T2V generation, while facilitating reproducible research and the development of more robust and secure video generation systems.