VANS-Data-100K
收藏资源简介:
VANS-Data-100K是由香港城市大学与快手科技联合构建的专用视频数据集,旨在支持视频下一事件预测任务的研究。该数据集包含10万条高质量样本,其中3万为流程性任务数据,7万为预测性场景数据,每条样本由输入视频、问题及多模态答案构成。数据集通过四阶段流程精心构建:从COIN、YouCook2等来源采集原始视频,经镜头分割、片段筛选后,利用大语言模型生成多样化问答对与推理链。该数据集主要应用于视频事件推理与生成领域,致力于解决从文本描述到动态视频演示的跨模态对齐难题。
VANS-Data-100K is a specialized video dataset jointly constructed by City University of Hong Kong and Kuaishou Technology, aiming to support research on video next-event prediction tasks. This dataset contains 100,000 high-quality samples, among which 30,000 are procedural task data and 70,000 are predictive scenario data. Each sample consists of an input video, a question, and multimodal answers. The dataset is meticulously constructed via a four-stage pipeline: raw videos are collected from sources such as COIN and YouCook2, followed by shot segmentation and clip filtering, then diverse question-answer pairs and reasoning chains are generated using Large Language Models (LLMs). This dataset is primarily applied in the field of video event reasoning and generation, aiming to address the cross-modal alignment challenges between text descriptions and dynamic video demonstrations.
VANS 数据集概述
数据集基本信息
- 项目名称:Video-as-Answer: Predict and Generate Next Video Event with Joint-GRPO
- 项目地址:https://github.com/KlingTeam/VANS
- 模型地址:https://huggingface.co/KlingTeam/VANS
- 论文地址:https://arxiv.org/abs/2511.16669
- 项目网站:https://video-as-answer.github.io/
研究背景
- 核心任务:视频下一事件预测(Video-Next-Event Prediction, VNEP)
- 研究目标:将基于文本的下一事件预测扩展到动态视频响应
- 方法创新:从"讲述"转向"展示",为程序性学习和创意探索提供更直观和定制化的答案
技术方法
VANS架构
- 双路径处理:
- VLM路径:执行基于指令的推理,生成文本描述
- VDM路径:基于语义描述和视觉上下文合成视频
Joint-GRPO方法
- 两阶段强化学习:
- 阶段1:可视化友好的VLM调优 - 优化描述以实现视觉合理性
- 阶段2:上下文忠实的VDM适配 - 确保语义对齐和视觉连贯性
应用场景
程序性教学
- 烘焙烹饪指导
- 手工制作教学
- 创意活动指导
多未来预测
- 基于不同问题的多样化未来预测
- 现实场景与夸张电影场景的对比
- 情感反应的多样性展示
模型组件
- 基础模型:
- Qwen2.5-VL-3B(视觉语言模型)
- Wan2.1-T2V-1.3B(视频扩散模型)
- 完整模型:VANS模型
发布计划
- [ ] 发布VANS-Data-100K数据集
- [ ] 发布VANS模型
- [ ] 发布训练代码
- [x] 发布推理代码
- [x] 发布论文
引用信息
bibtex @misc{cheng2025videoasanswerpredictgeneratevideo, title={Video-as-Answer: Predict and Generate Next Video Event with Joint-GRPO}, author={Junhao Cheng and Liang Hou and Xin Tao and Jing Liao}, year={2025}, eprint={2511.16669}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2511.16669}, }

- 1Video-as-Answer: Predict and Generate Next Video Event with Joint-GRPO香港城市大学, 快手科技Kling团队 · 2025年



