README.md

---
license: cc-by-nc-4.0
pretty_name: ViewBench
tags:
- video
- world-models
- camera-conditioned-video
- view-consistency
- loop-closure
- ue5
- arxiv:2602.07854
---

# ViewBench

ViewBench is a dataset for camera-conditioned long-horizon video world models. It is designed to evaluate view consistency and loop closure: when a camera returns to a previously observed viewpoint, the generated observation should preserve stable scene structure and appearance.

This dataset accompanies the paper [Consistent Video World Model With Geometry-Aware Rotary Position Embedding](https://arxiv.org/abs/2602.07854).

Dataset mirrors:

- ModelScope: https://modelscope.cn/datasets/JEdward/viewbench-dataset
- Hugging Face: https://huggingface.co/datasets/JEdward/viewbench-dataset

## Highlights

- Complete yaw, pitch, and roll coverage for controlled camera motion.
- Round-trip loop-closure trajectories where the camera returns to previously visited viewpoints.
- 10 photorealistic UE5 environments spanning indoor, outdoor, urban, industrial, historical, and suburban scenes.
- Per-frame SE(3) camera-to-world poses and depth-based geometric overlap annotations.

## Dataset Contents

This v1 release contains the public ViewBench training split described in the paper: 1,059 UE5-rendered video sequences, about 500k frames at 30 fps, across 10 photorealistic environments.

The release is organized into two trajectory groups:

- `pure_rotation`: stationary-camera rotate-away-rotate-back trajectories for loop closure.
- `rotation_translation`: compact exploration trajectories with both rotation and translation.

The original internal directories were `STAGE1` and `STAGE3`. In this public release they are renamed to `pure_rotation` and `rotation_translation`.

The paper's held-out evaluation set is separately collected and is not included in this training release unless explicitly provided in a later update.

## Files

The dataset is distributed as `tar.zst` shards plus `manifest.json`:

- `pure_rotation_0000.tar.zst` ... `pure_rotation_0011.tar.zst` (600 sequences)
- `rotation_translation_0000.tar.zst` ... `rotation_translation_0009.tar.zst` (459 sequences)
- `manifest.json`

`manifest.json` records the shard membership, sequence IDs, original-to-public stage mapping, and archive contents.

Each archive extracts into:

```text
ViewBench4Training/
  pure_rotation/
    frames/{sequence_id}/
    jsons/{sequence_id}.json
    metadata/{sequence_id}/
  rotation_translation/
    frames/{sequence_id}/
    jsons/{sequence_id}.json
    metadata/{sequence_id}/
```

## Data Format

Each sequence contains:

- EXR frames with RGB/depth information.
- Per-frame camera poses in `jsons/{sequence_id}.json`.
- Raw metadata in `metadata/{sequence_id}/tickStatus.jsonl` where available.
- Depth-based frame overlap labels in `metadata/{sequence_id}/overlap.json` where available.

Camera convention:

- UE left-handed coordinates: `X=forward`, `Y=right`, `Z=up`.
- Position is measured in centimeters.
- Rotation is `[pitch, roll, yaw]` in degrees.
- `c2w` is a 4x4 camera-to-world SE(3) matrix.
- Rotation convention: `R = Rz(yaw) * Ry(pitch) * Rx(roll)`.

## Usage

Download from ModelScope:

```bash
modelscope download --dataset JEdward/viewbench-dataset --local_dir ViewBench-v1
```

After all shards are downloaded, extract them into a single directory:

```bash
mkdir -p ViewBench4Training
for shard in ViewBench-v1/pure_rotation_*.tar.zst ViewBench-v1/rotation_translation_*.tar.zst; do
  tar --zstd -xf "$shard" -C ViewBench4Training
done
```

Repeat extraction for all shards listed in `manifest.json`.

## License

This dataset is released for non-commercial research use under CC BY-NC 4.0-style terms.

Users may use, copy, and redistribute the dataset for academic and non-commercial research purposes, provided that they give appropriate attribution and cite the accompanying paper.

Commercial use, resale, or redistribution as part of a commercial dataset or product is not permitted without prior written permission from the authors.

The dataset contains UE5-rendered outputs from third-party scene assets. The release does not include raw UE assets, source asset files, or engine content. Users are responsible for ensuring that their downstream use complies with applicable third-party asset terms.

## Citation

```bibtex
@inproceedings{
xiang2026consistent,
title={Consistent Video World Model With Geometry-Aware Rotary Position Embedding},
author={Chendong Xiang and Jiajun Liu and Jintao Zhang and Xiao Yang and Zhengwei Fang and Shizun Wang and Zijun Wang and Yingtian Zou and Hang Su and Jun Zhu},
booktitle={ICLR 2026 the 2nd Workshop on World Models: Understanding, Modelling and Scaling},
year={2026},
url={https://openreview.net/forum?id=eXgmwOOvlR}
}
```