PlayWorld
收藏资源简介:
PlayWorld是由香港大学、快手科技Kling团队等机构联合构建的基准数据集,旨在评估视频世界模型在长时交互目标下的性能。该数据集包含171个人工标注的场景,每个场景设定一个长时目标(如360度转身或走入水中),并配套基础动作序列,覆盖几何一致性、交互保真度、视线外演化与洞察演化四个核心维度,同时纳入基本能力指标。数据构建过程从Pexels和Google Images采集多样化初始图像,由标注者定义目标、编写动作序列并生成超过820个VQA问题,最终产出1400余条交互视频。该数据集解决了现有基准因固定动作预设无法公平比较不同模型长时交互能力的问题,实验揭示当前模型在空间一致性与状态持久演化方面仍存在显著不足。
PlayWorld is a benchmark dataset jointly constructed by The University of Hong Kong, the Kling Team of Kwai Technology and other institutions, designed to evaluate the performance of video world models under long-term interactive goals. This dataset contains 171 manually annotated scenarios, each with a long-term objective such as a 360-degree turn or walking into water, paired with basic action sequences. It covers four core evaluation dimensions: geometric consistency, interaction fidelity, off-screen evolution and insight-driven evolution, as well as basic capability metrics. During the dataset construction process, diverse initial images were collected from Pexels and Google Images. Annotators defined the objectives, drafted the action sequences and generated over 820 VQA questions, ultimately producing more than 1,400 interactive videos. This dataset addresses the issue that existing benchmarks cannot fairly compare the long-term interactive capabilities of different models due to their fixed action presets. Experimental results reveal that current models still have significant shortcomings in spatial consistency and persistent state evolution.
PlayWorld 数据集详情
简介
PlayWorld 是一个用于基准测试交互式世界模型的评估数据集,通过智能体玩家在长周期目标下的表现来评估世界模型能力。
核心特征
- 评估方式:无人工操作,全自动评估,引擎基于 Web 的世界模型执行,自动进入初始世界状态、等待世界可交互、派发控制指令并捕获观察结果,最后保留最终视频与执行记录。
- 数据规模:171 个案例(Cases),1,417 个视频(Videos),每次推演(Rollout)时长 10–60 秒,覆盖 9 个以上模型。
- 模型来源:涵盖多个代表性模型,如 Alibaba 的 HappyOyster、Google DeepMind 的 Genie 3、Tencent 的 HY-World2 等。
数据集内容
- 精选模型推演(Model Demos):展示各模型在多种任务下的推演视频,任务类型包括几何一致性(Geometry Consistency)、交互保真度(Interaction Fidelity)等,每个视频标注了任务族(Task Family)与推演时长。
- 完整玩家过程(Full Player Process):提供从世界创建、自适应动作执行到最终结果保存的完整记录视频,展现了不同模型在实际任务中的完整执行流程。
评估维度与对比
PlayWorld 在多个关键维度上覆盖完整,相比其他基准具有更全面的评估能力:
| 维度 | PlayWorld 覆盖情况 |
|---|---|
| 输入形式 | 文本 + 图像 + 目标 |
| 闭环自适应(Closed-loop adaptation) | ✓ |
| 长周期重访(Long-horizon revisit) | ✓ |
| 几何一致性(Geometry consistency) | ✓ |
| 状态演化(State evolution) | ✓ |
| 交互保真度(Interaction fidelity) | ✓ |
| 时间尺度/视频 | 10–60 秒 |
访问链接
- 项目主页:https://kxding.github.io/project/PlayWorld/
- 论文(Paper)、GitHub、排行榜(Leaderboard)及数据集(Dataset)相关资源均已提供。




