---
license: apache-2.0
task_categories:
- video-text-to-text
---

<p align="center">
  <img src="assets/logo2.png" alt="FramePack-U logo" style="width: 65%;" />
</p>

<h1 align="center">
  Elephant in the Fridge: Constant-Memory Frame Packing for Long Video Understanding
</h1>

<p align="center">
  <a href="https://zeld0n.github.io/shuo_homepage/">Shuo Gao</a>, <a href="https://hellomuffin.github.io/">Chenhao Zheng</a>, <a href="https://jason718.github.io/">Zhongzheng Ren</a>.
</p>
<p align="center">
<a href="https://zeld0n.github.io/framepack-U/"><img src="https://img.shields.io/badge/Homepage-FramePackU-green" alt="Homepage"></a>
<a href="https://arxiv.org/abs/2610.16118"><img src="https://img.shields.io/badge/arXiv-FramePackU-red?logo=arxiv" alt="GitHub"></a>
<a href="https://github.com/ZelD0n/framepack-U"><img src="https://img.shields.io/badge/GitHub-FramePackU-blue?logo=github" alt="GitHub"></a>
<a href="https://huggingface.co/datasets/ZelD0n/FramePack-U"><img src="https://img.shields.io/badge/HuggingFace-FramePackU-yellow?logo=huggingface" alt="Hugging Face"></a>
<a href="https://www.modelscope.cn/datasets/ZelD0n/FramePack-U"><img src="https://img.shields.io/badge/ModelScope-FramePackU-purple?logo=modelscope" alt="ModelScope"></a>
</p>

<p align="center">
  <img src="assets/framepack_framework.png" alt="FramePack-U framework" style="width: 80%;"/>
</p>

## What does this repo contain?
This repo contains:
- **Offline CLIP scores**: Relevance between each frame and text query, calculated by CLIP-ViT-B/32.
- **Offline frame feature**: Visual feature extracted by CLIP-ViT.

## How to use it?
To avoid massive online calculating, you can download these data and put these offline scores and feature in the `local/` directory of the repo, or download wherever you want and link them to this directory. 

Unzip the `clip_scores.zip` and `frame_feature.zip` file by this script:
```
unzip /path/to/clip_scores.zip
unzip /path/to/frame_feature.zip
```

Now you can run the evaluation on LongVideoBench, MLVU and VideoMME with their pre-calculated frame feature.

## Citation
If you find this work helpful, please consider citing our paper:
```bibtex
@inproceedings{gao2026elephant,
  title = {Elephant in the Fridge: Constant-Memory Frame Packing for Long Video Understanding},
  author = {Gao, Shuo and Zheng, Chenhao and Ren, Zhongzheng},
  booktitle = {Neural Information Processing Systems (NeurIPS)},
  year = {2026}
}
```