T2VTextBench
收藏资源简介:
T2VTextBench是一个用于评估现代文本到视频模型中文本操作的全面基准。该数据集包含73个提示,旨在评估文本到视频模型在复杂时间动态下生成文本的能力。每个提示都设计成评估模型在现实世界场景中的文本操作和上下文一致性。数据集分为六个类别:逐步或符号可视化、应用程序和Web UI模拟、日常数字时刻、电影或演示场景、数学相关和多语言(中文)。
T2VTextBench is a comprehensive benchmark for evaluating text manipulation in modern text-to-video models. This dataset comprises 73 prompts intended to assess the capability of text-to-video models to perform text manipulation under complex temporal dynamics. Each prompt is designed to evaluate a model's text manipulation performance and contextual consistency within real-world scenarios. The dataset is categorized into six groups: step-by-step or symbolic visualization, application and Web UI simulation, daily digital moments, movie or demo scenarios, mathematics-related content, and multilingual (Chinese) content.

- 1T2VTextBench: A Human Evaluation Benchmark for Textual Control in Video Generation Models桂林电子科技大学, 亚利桑那大学, 威斯康星大学麦迪逊分校, 加州大学伯克利分校, 亚利桑那州立大学 · 2025年



