T2VCountBench
收藏资源简介:
T2VCountBench是一个专门针对评估现代文本到视频生成模型计数能力的基准。该数据集由10种不同类型的文本到视频生成模型产生的视频组成,使用统一的人类评估协议来确保评估的准确性和公正性。数据集包含165个独特的文本提示,覆盖了人类、自然和人工三种对象类别,并考虑了场景转换、运动约束和风格约束。该数据集旨在解决文本到视频生成模型在遵循基本数值约束方面的挑战。
T2VCountBench is a benchmark specifically designed to evaluate the counting capabilities of modern text-to-video generation models. The dataset consists of videos generated by 10 different types of text-to-video generation models, and a unified human evaluation protocol is employed to guarantee the accuracy and fairness of the evaluation. It includes 165 distinct text prompts covering three object categories: humans, natural objects, and artificial objects, while also considering scene transitions, motion constraints and style constraints. This benchmark is intended to address the challenges that text-to-video generation models encounter when adhering to basic numerical constraints.

- 1Can You Count to Nine? A Human Evaluation Benchmark for Counting Limits in Modern Text-to-Video Models桂林电子科技大学, 俄亥俄州立大学, 亚利桑那大学, 香港大学, 威斯康星大学麦迪逊分校, 加州大学伯克利分校, 独立研究者 · 2025年



