KO | EN

GitLite

Minimalist Open Source Repository Search Engine

gitlite — search

#dsh #dsh-plugin #deepseek-harness #typescript #ai-agent #codex #developer-tools #claude-code #windows #mcp #python #desktop-app

VANS

★ 119 Open GitHub ↗

[CVPR 2026] Video-as-Answer: Predict and Generate Next Video Event with Joint-GRPO

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

<h1 align="center">Video-as-Answer: Predict and Generate Next Video Event with Joint-GRPO</h1>

[Junhao Cheng<sup>1†</sup>](https://donahowe.github.io/),
[Liang Hou<sup>2</sup>](https://liang-hou.github.io/),
[Xin Tao<sup>2</sup>](https://www.xtao.website/),
[Jing Liao<sup>1</sup>](https://scholar.google.com/citations?user=3s9f9VIAAAAJ&hl=en)  
<sup>1</sup>City University of Hong Kong  <sup>2</sup>Kling Team, Kuaishou Technology  
<sup>†</sup> This work was conducted during the author's internship at Kling Team, Kuaishou Technology

</div>

## 🔎 Introduction

We pioneer **Video-Next-Event Prediction (VNEP)**, extending text-based next-event prediction to dynamic video responses. This shift from *telling* to *showing* enables more intuitive and customized answers for procedural learning and creative exploration.

To tackle VNEP, we propose **VANS**, a model that aligns a Vision-Language Model (VLM) with a Video Diffusion Model (VDM) through our **Joint-GRPO** post-training approach. Our method bridges the semantic-to-visual gap of VLM and VDM, enabling high-quality video event prediction and generation.

## 🏗️ Method

<div align="center">
  <table>
    <tr>
      <td align="center" width="33%">
        <img src="assets/images/model.png" alt="VANS Architecture" style="max-width: 100%; height: auto; max-height: 300px;">
        <br>
        <em>VANS Architecture: Dual-path processing with VLM for reasoning and VDM for generation</em>
      </td>
      <td align="center" width="45%">
        <img src="assets/images/grpo.png" alt="Joint-GRPO" style="max-width: 100%; height: auto; max-height: 300px;">
        <br>
        <em>Joint-GRPO: Two-stage co-steering optimization</em>
      </td>
    </tr>
  </table>
</div>

### Key Components

**VANS Architecture**: Processes input videos and questions through dual pathways:
- **VLM Path**: Performs instruction-grounded reasoning to generate textual captions
- **VDM Path**: Synthesizes videos conditioned on semantic captions and visual context

**Joint-GRPO**: Our two-stage reinforcement learning approach:
- **Stage 1**: Visualization-friendly VLM tuning - optimizes captions for visual plausibility
- **Stage 2**: Context-faithful VDM adaptation - ensures semantic alignment and visual coherence

## 🚩 Plan
- [ ] Release training codes
- [x] Release VANS-Data-100K [construction pipelines](https://github.com/KlingAIResearch/VANS/tree/main/data_pipeline)
- [x] Release VANS model [checkpoints](https://huggingface.co/KlingTeam/VANS/tree/main)
- [x] Release inference codes
- [x] Release [paper](https://arxiv.org/abs/2511.16669)

## 🎬 Results

### 🍳 Procedural Teaching

| Case | Input Video | Question | VANS Output |
|:----:|:-----------:|:--------:|:----------:|
| 1 | <img src="assets/videos/proce_1/input.gif" width="220" height="220" style="border-radius:6px;" alt="Input Video 1"> | "Show me the next step for baked chicken Parmesan." | <img src="assets/videos/proce_1/jointgrpo.gif" width="220" height="220" style="border-radius:6px;" alt="Output Video 1"> |
| 2 | <img src="assets/videos/proce_2/input.gif" width="220" height="220" style="border-radius:6px;" alt="Input Video 2"> | "Hi, I want to make slime. What should I do next?" | <img src="assets/videos/proce_2/jointgrpo.gif" width="220" height="220" style="border-radius:6px;" alt="Output Video 2"> |
| 3 | <img src="assets/videos/proce_3/input.gif" width="220" height="220" style="border-radius:6px;" alt="Input Video 3"> | "Hey AI assistant, I'm making a paper windmill and just uploaded a video. What should I do next?" | <img src="assets/videos/proce_3/jointgrpo.gif" width="220" height="220" style="border-radius:6px;" alt="Output Video 3"> |

</div>

### 🔮 Multi-Future Prediction

**Same input video, different questions lead to diverse future predictions:**

<table>
<tr>
<td align="center" colspan="3"><strong>Input Video</strong></td>
</tr>
<tr>
<td align="center" colspan="3">
<img src="assets/videos/gen_1/input.gif" width="220" height="220" style="border-radius:8px;" alt="Kitchen Input">
</td>
</tr>
<tr>
<td align="center">
<img src="assets/videos/gen_1/con1.gif" width="220" height="220" style="border-radius:6px;" alt="Realistic Reaction"><br>
<small>"What if she gets burned in her daily life?"</small>
</td>
<td align="center">
<img src="assets/videos/gen_1/con2.gif" width="220" height="220" style="border-radius:6px;" alt="Dramatic Reaction"><br>
<small>"What if she gets burned in an exaggerated movie?"</small>
</td>
<td align="center">
<img src="assets/videos/gen_1/con3.gif" width="220" height="220" style="border-radius:6px;" alt="Comedic Reaction"><br>
<small>"What if she eats something spicy in an exaggerated movie?"</small>
</td>
</tr>
</table>

<table>
<tr>
<td align="center" colspan="3"><strong>Input Video</strong></td>
</tr>
<tr>
<td align="center" colspan="3">
<img src="assets/videos/gen_2/input.gif" width="220" height="220" style="border-radius:8px;" alt="Emotional Input">
</td>
</tr>
<tr>
<td align="center">
<img src="assets/videos/gen_2/con1.gif" width="220" height="220" style="border-radius:6px;" alt="Grandson Reaction"><br>
<small>"Show her reaction if she sees her grandson."</small>
</td>
<td align="center">
<img src="assets/videos/gen_2/con2.gif" width="220" height="220" style="border-radius:6px;" alt="Husband Reaction"><br>
<small>"Show her reaction if she sees her husband."</small>
</td>
<td align="center">
<img src="assets/videos/gen_2/con3.gif" width="220" height="220" style="border-radius:6px;" alt="Death Reaction"><br>
<small>"Show her reaction if she sees the personification of death."</small>
</td>
</tr>
</table>

</div>

## 🚀 Quick Start

### 🎯 Environment Setup

To set up the environment for inference, you can run the following command:
```shell
git clone https://github.com/KlingTeam/VANS.git
cd VANS

conda create -n VANS python==3.12 -y
conda activate VANS

pip install requirements.txt
cd vans/models_mllm/qwen-vl-utils
pip install -e .[decord]
cd ...
```

### 🌎 Download Models

To get started, download the VANS base models:
- **[Qwen2.5-VL-3B](https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct)** - The Vision-Language Model
- **[Wan2.1-T2V-1.3B](https://huggingface.co/Wan-AI/Wan2.1-T2V-1.3B)** - The Video Diffusion Model

Then download the complete VANS model:  
**[VANS Model Download](https://huggingface.co/KlingTeam/VANS)**

### 🧸 Demo
To run local gradio demo:
```shell
python app.py
```

## 📜 Citation

If you find our work helpful, please consider giving a star 🌟 and citation 📝

```BibTeXw
@article{cheng2025video,
  title={Video-as-Answer: Predict and Generate Next Video Event with Joint-GRPO},
  author={Cheng, Junhao and Hou, Liang and Tao, Xin and Liao, Jing},
  journal={arXiv preprint arXiv:2511.16669},
  year={2025}
}
```

VANS

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README