README.md

---
license: cc-by-4.0
task_categories:
- video-classification
- text-to-video
language:
- en
tags:
- causality
- arrow-of-time
- video-generation
- video
- benchmark
- world-model
- violation-of-expectation
- diffusion-models
pretty_name: YoCausal
source_datasets:
- extended
size_categories:
- 1K<n<10K
---

<p align="center">
  <img src="asset/Logo.png" alt="YoCausal Logo" width="400">
</p>

# YoCausal: A Causality Benchmark for Video Generation Models

<p align="center">
  <a href="https://www.youzhexie.me/">You-Zhe Xie</a><sup>🦊🌐*</sup>,
  <a href="https://www.yhlizzz.com/">Yu-Hsuan Li</a><sup>🦊*</sup>,
  <a href="https://jayinnn.dev/">Jie-Ying Lee</a><sup>🦊</sup>,
  <a href="https://kpzhang93.github.io/">Kaipeng Zhang</a><sup>🌐</sup>,<br>
  <a href="https://yulunalexliu.github.io/">Yu-Lun Liu</a><sup>🦊&dagger;</sup>,
  <a href="https://lightchaserx.github.io/">Zhixiang Wang</a><sup>🌐&dagger;</sup>
</p>
<p align="center">
  🦊 National Yang Ming Chiao Tung University &nbsp;&nbsp; 🌐 Shanda AI Research Tokyo<br>
  <sup>*</sup> Equal contribution &nbsp;&nbsp; <sup>&dagger;</sup> Corresponding authors
</p>

## Overview

**YoCausal** (ようこそ, *Yōkoso*) evaluates whether video generation models are sensitive to causal temporal structure. Each natural video is paired with its time-reversed counterpart; models are compared on forward versus backward denoising loss using the same prompt and noise condition.

YoCausal accompanies *YoCausal: How Far is Video Generation from World Model? A Causality Perspective* and is intended for benchmark evaluation rather than training.

- **Project page:** https://www.youzhexie.me/papers/YoCausal/index.html
- **Huggingface paper page:** https://huggingface.co/papers/2605.30346
- **Arxiv Paper page:** http://arxiv.org/abs/2605.30346
- **Dataset repository:** https://huggingface.co/datasets/YouZhe/YoCausal-dataset
- **Code repository:** https://github.com/youzhe0305/YoCausal
- **Contents:** 1,232 forward-backward pairs (2,464 MP4 files) with English prompts and JSON metadata

## Dataset Structure

| Subset | Source collection | Pairs | Duration | Content |
|---|---|---:|---|---|
| `general` | Moments in Time | 500 | 3 s | Everyday events and actions |
| `physics` | Physics-IQ | 132 | 5 s | Physical phenomena |
| `human` | Kinetics-400 subset | 400 | 3 s | Human actions |
| `animal` | Animal Kingdom | 200 | 3 s | Animal behavior |
| **Total** |  | **1,232** |  |  |

```text
subset/<name>/
  dataset_metadata.json
  fwd/*.mp4
  bwd/*.mp4
```

Each metadata record refers to one pair. `fwd/` contains natural videos and `bwd/` contains the corresponding temporally reversed videos.

## Metadata

Each `dataset_metadata.json` is a JSON array with the following fields:

| Field | Description |
|---|---|
| `id` | Entry identifier within a subset. |
| `video_path_forward`, `video_path_backward` | Preparation-time video paths; use their basenames to locate released MP4 files. |
| `prompt` | Caption used for both temporal directions during evaluation. |
| `dataset_source`, `category` | Source collection and content category. |
| `meta` | Recorded `fps`, `[height, width]` resolution, and `total_frames`. |
| `vlm_causality` | VLM-based binary label for observable causal structure. |
| `human_discriminable` | Human temporal-direction judgment: `true`, `false`, or `"unknown"`(see paper for details). |

## Loading

Videos are stored with Git LFS:

```bash
git lfs install
git clone https://huggingface.co/datasets/YouZhe/YoCausal-dataset
```

The metadata preserves preparation-time paths. Resolve released videos by filename:

```python
import json
from pathlib import Path

root = Path("YoCausal-dataset")
subset = "animal"
with (root / "subset" / subset / "dataset_metadata.json").open() as f:
    sample = json.load(f)[0]

name = Path(sample["video_path_forward"]).name
fwd = root / "subset" / subset / "fwd" / name
bwd = root / "subset" / subset / "bwd" / name
```

## Evaluation Protocol

For each pair, the model receives the same prompt and matched noise for both directions. A forward win occurs when the loss on the natural video is lower than on its reversed counterpart.

| Metric | Meaning |
|---|---|
| **RSI** | Forward win rate, measuring temporal-direction sensitivity. |
| **RSI(Dc)** / **RSI(Dnc)** | RSI on `vlm_causality=true` / `false` examples. |
| **CCI** | `RSI(Dc) - RSI(Dnc)`, intended to separate causal sensitivity from general temporal preference. |

When evaluating human-discriminable versus non-discriminable subsets, entries with `human_discriminable="unknown"` should be retained as unlabeled for that split and excluded from the `true` / `false` subgroup comparison.

## Citation

```bibtex
@article{xie2026yocausal,
  title   = {YoCausal: How Far is Video Generation from World Model? A Causality Perspective},
  author  = {Xie, You-Zhe and Li, Yu-Hsuan and Lee, Jie-Ying and Zhang, Kaipeng and Liu, Yu-Lun and Wang, Zhixiang},
  journal = {arXiv preprint arXiv:2605.30346},
  year    = {2026}
}
```