---
pretty_name: UniBench
license: apache-2.0
language:
- en
size_categories:
- 10K<n<100K
tags:
- video-generation
- benchmark
- depth-estimation
- optical-flow
- multimodal
- world-aware
arxiv: "2512.07831"
configs:
- config_name: uedata
  data_files:
  - split: train
    path: UniBench/UEData/train/train.csv
  - split: validation
    path: UniBench/UEData/eval/eval.csv
- config_name: realdata
  data_files:
  - split: test
    path: UniBench/RealData/eval.csv
---

# UniBench

**The official evaluation benchmark for
[UnityVideo: Unified Multi-Modal Multi-Task Learning for Enhancing World-Aware Video Generation](https://arxiv.org/abs/2512.07831).**

UniBench evaluates world-aware video generation and estimation across RGB video,
optical flow, and depth. This repository contains the benchmark data and metadata
used by the paper; it is not the large-scale training-data release.

## UniBench vs. OpenUni

| Repository | Purpose | Contents |
| --- | --- | --- |
| **[KlingTeam/UniBench](https://huggingface.co/datasets/KlingTeam/UniBench)** | Evaluation benchmark reported in the UnityVideo paper | UEData RGB/RAFT/depth triplets and a public real-video evaluation subset |
| **[JackAILab/OpenUni](https://huggingface.co/datasets/JackAILab/OpenUni)** | Large-scale training data for UnityVideo | The OpenUni training corpus and its multimodal annotations |

Use **OpenUni to train or fine-tune models** and **UniBench to run the paper's
benchmark evaluation**. Keeping the two repositories separate avoids mixing
training samples with benchmark splits.

## Dataset Contents

| Subset | Split | Cases | Modalities | Storage |
| --- | --- | ---: | --- | --- |
| UEData | `train` / reference | 29,800 | RGB, RAFT optical flow, depth | 100 uncompressed tar shards |
| UEData | `eval` | 200 | RGB, RAFT optical flow, depth | 1 uncompressed tar shard |
| RealData | `eval` | 100 | RGB video | Directly browsable MP4 files |

UEData contains 30,000 cases in total. Each UEData row points to three aligned
videos: the source RGB video (`ceph_path`), optical flow visualization (`raft`),
and depth visualization (`depth`). RealData contains 100 public Koala36M samples
and does not provide RAFT or depth files.

```text
UniBench/
├── manifest.json
├── UEData/
│   ├── train/
│   │   ├── train.csv
│   │   ├── shard_manifest.csv
│   │   └── shards/train-00000.tar ... train-00099.tar
│   └── eval/
│       ├── eval.csv
│       ├── shard_manifest.csv
│       └── shards/eval-00000.tar
└── RealData/
    ├── eval.csv
    └── videos/*.mp4
```

All paths in the CSV files are relative to their split directory. Extracting a
UEData shard inside its split directory restores the referenced `videos/...`
paths. Every tar shard has a neighboring `.sha256` checksum file, and
`shard_manifest.csv` maps every video path to its shard.

## Quick Start

Install the required clients:

```bash
pip install -U huggingface_hub datasets pandas
```

### Read Metadata Only

The Hugging Face `datasets` integration loads the CSV metadata without
downloading the large video shards:

```python
from datasets import load_dataset

ue = load_dataset("KlingTeam/UniBench", "uedata")
real = load_dataset("KlingTeam/UniBench", "realdata")

print(ue["train"][0])       # UEData reference/train metadata
print(ue["validation"][0])  # UEData evaluation metadata
print(real["test"][0])      # RealData evaluation metadata
```

### Download Evaluation Data

Download only the held-out UEData split and the public real-video subset:

```bash
hf download KlingTeam/UniBench \
  --repo-type dataset \
  --include "UniBench/UEData/eval/**" \
  --include "UniBench/RealData/**" \
  --include "UniBench/manifest.json" \
  --local-dir ./unibench
```

Verify and extract the UEData evaluation shard:

```bash
cd ./unibench/UniBench/UEData/eval
(cd shards && sha256sum -c eval-00000.tar.sha256)
tar -xf shards/eval-00000.tar
```

After extraction, paths such as `videos/example.mp4` in `eval.csv` resolve
relative to `./unibench/UniBench/UEData/eval/`.

### Download the Full Benchmark

The full release is approximately 375 GiB. Make sure the destination has enough
free space for both the downloaded tar files and extracted videos.

```bash
hf download KlingTeam/UniBench \
  --repo-type dataset \
  --local-dir ./unibench
```

Extract all UEData shards:

```bash
cd ./unibench/UniBench/UEData/train
for shard in shards/*.tar; do tar -xf "$shard"; done

cd ../eval
for shard in shards/*.tar; do tar -xf "$shard"; done
```

To save space, use `shard_manifest.csv` to identify and download only the shards
needed for selected cases.

## Metadata Format

| Column | Description |
| --- | --- |
| `ceph_path` | Relative path to the RGB video |
| `caption_list` | JSON-encoded list of text captions |
| `raft` | Relative path to the optical-flow video; empty for RealData |
| `depth` | Relative path to the depth video; empty for RealData |
| `duration` | Duration in seconds |
| `fps` | Frames per second |
| `height` | Video height in pixels |
| `width` | Video width in pixels |

Example for resolving one extracted UEData case:

```python
import json
from pathlib import Path

import pandas as pd

split_dir = Path("./unibench/UniBench/UEData/eval")
row = pd.read_csv(split_dir / "eval.csv").iloc[0]

sample = {
    "rgb": split_dir / row["ceph_path"],
    "raft": split_dir / row["raft"],
    "depth": split_dir / row["depth"],
    "caption": json.loads(row["caption_list"])[0],
}
print(sample)
```

## Citation

If you use UniBench or OpenUni, please cite:

```bibtex
@article{huang2025unityvideo,
  title={UnityVideo: Unified Multi-Modal Multi-Task Learning for Enhancing World-Aware Video Generation},
  author={Huang, Jiehui and Zhang, Yuechen and He, Xu and Gao, Yuan and Cen, Zhi and Xia, Bin and Zhou, Yan and Tao, Xin and Wan, Pengfei and Jia, Jiaya},
  journal={arXiv preprint arXiv:2512.07831},
  year={2025}
}
```

## License and Source Data

This repository is released under the Apache 2.0 license. RealData samples are
drawn from the public Koala36M dataset. Users are also responsible for complying
with the terms of any applicable upstream datasets and for using the benchmark
for lawful research purposes.