T2VEval-Bench
收藏资源简介:
T2VEval-Bench是由中国传媒大学信息与通信工程学院构建的大规模文本生成视频评估基准数据集。该数据集包含148个文本提示和1783个由12个不同模型生成的视频,涵盖了人类、动物、景观和想象等四个主题。数据集通过主观评估收集了五个关键评分维度:整体印象、视频质量、美学质量、真实性和文本-视频一致性。数据集的创建过程包括文本提示生成、视频生成和主观评分收集,旨在解决文本生成视频质量评估中的复杂失真问题。该数据集的应用领域主要集中在文本到视频生成模型的优化和质量评估,为未来的研究提供了可重复且稳健的基准。
T2VEval-Bench is a large-scale text-to-video generation evaluation benchmark dataset constructed by the School of Information and Communication Engineering, Communication University of China. This dataset includes 148 text prompts and 1,783 videos generated by 12 distinct models, covering four themes: humans, animals, landscapes, and imaginative content. Subjective evaluations were conducted to collect five key scoring dimensions: overall impression, video quality, aesthetic quality, authenticity, and text-video consistency. The dataset creation process consists of text prompt generation, video generation and subjective scoring collection, aiming to resolve the complex distortion issues in the quality evaluation of text-to-video generation models. The main application scenarios of this dataset focus on the optimization and quality assessment of text-to-video generation models, providing a reproducible and robust benchmark for future research.

- 1Comprehensive Subjective and Objective Evaluation Method for Text-generated Video中国传媒大学信息与通信工程学院 · 2025年



