genintel/CRONOS-benchmark
收藏资源简介：
CRONOS-Benchmark 是一个受控的合成基准数据集，用于评估视频世界模型中的反事实物理一致性。它测试生成视频模型是否能在不同场景、对象和视角下正确模拟三种基本物理事件类型——物体下落、物体碰撞和物体遮挡。数据集包含675个序列，在受控模拟环境中渲染，每个序列捕获其中一种物理事件类型，并提供真实的多模态注释（RGB、深度、掩码），适用于任何接受图像或视频条件输入的视频生成模型的标准化评估。
CRONOS-Benchmark is a controlled synthetic benchmark dataset for evaluating counterfactual physical consistency in video world models. It tests whether video generation models can correctly simulate three fundamental types of physical events—object falling, object collision, and object occlusion—across diverse scenarios, objects and viewpoints. The dataset contains 675 sequences rendered in a controlled simulation environment, with each sequence capturing one of the three physical event types. It also provides authentic multimodal annotations (RGB, depth, masks) to enable standardized evaluation for any video generation model that accepts image or video conditional inputs.
