StateBench
收藏资源简介:
StateBench是一个用于评估视频续写中世界状态推理能力的基准数据集,源自论文《Do Video Generators Track the World Across Segments? A Benchmark and Method for World-State Reasoning in Video Continuation》。该数据集包含200个跨片段续写任务,分为三个难度级别:past_visible(85个任务,目标对象在先前帧中可见,测试时间记忆)、occluded_process(65个任务,状态变化发生在对象被遮挡时,测试过程推理)、complex_transition(50个任务,多个状态转换或冲突线索,测试组合推理)。每个任务包含元数据字段:id、difficulty、target_object、shots、expected_state、checklist和reference_frames。该数据集旨在评估视频生成模型在续写过程中对物体状态变化的理解与预测能力。
StateBench is a benchmark dataset for evaluating world-state reasoning capabilities in video continuation, derived from the paper Do Video Generators Track the World Across Segments? A Benchmark and Method for World-State Reasoning in Video Continuation. It contains 200 cross-segment continuation tasks across three difficulty levels: past_visible (85 tasks, where the target object is visible in previous frames, testing temporal memory), occluded_process (65 tasks, where state changes occur when the object is occluded, testing process reasoning), and complex_transition (50 tasks, involving multiple state transitions or conflicting cues, testing compositional reasoning). Each task includes metadata fields: id, difficulty, target_object, shots, expected_state, checklist, and reference_frames. The dataset aims to evaluate video generation models ability to understand and predict object state changes during continuation.
StateBench 数据集详情
数据集简介
StateBench是一个用于视频延续中的世界状态推理(World-State Reasoning in Video Continuation)的基准测试数据集,源自论文《Do Video Generators Track the World Across Segments? A Benchmark and Method for World-State Reasoning in Video Continuation》(论文链接:https://huggingface.co/papers/2609.03673)。
许可证
- Apache-2.0
任务类别
- 文本到视频(text-to-video)
数据集规模与难度分级
该基准包含200个跨片段延续任务,按难度分为三个等级:
| 难度级别 | 数量 | 描述 |
|---|---|---|
past_visible(过去可见) |
85 | 目标物体在先前的帧中可见——测试时间记忆能力 |
occluded_process(遮挡过程) |
65 | 状态变化发生在物体被遮挡期间——测试过程推理能力 |
complex_transition(复杂转换) |
50 | 多个状态转换或冲突线索——测试组合推理能力 |
数据字段
每个任务包含以下元数据:
- id:任务标识
- difficulty:难度等级
- target_object:目标物体
- shots:镜头/片段信息
- expected_state:预期状态
- checklist:检查清单
- reference_frames:参考帧
相关资源
- 代码仓库:https://github.com/AMAP-ML/StateAgent




