README.md

---
license: apache-2.0
task_categories:
  - video-to-video
  - image-to-video
tags:
  - benchmark
  - physics
  - causal-reasoning
  - video-generation
  - world-models
size_categories:
  - 10G<n<100G
---

# CRONOS-Benchmark dataset

[![Paper](https://img.shields.io/badge/Paper-arXiv-red)](https://arxiv.org/abs/2605.23699)
[![Project Page](https://img.shields.io/badge/Project-Page-blue)](https://genintel.github.io/CRONOS/)
[![Dataset](https://img.shields.io/badge/Dataset-HuggingFace-yellow)](https://huggingface.co/datasets/genintel/CRONOS-benchmark)
[![Code](https://img.shields.io/badge/Code-GitHub-black)](https://github.com/GenIntel/CRONOS-benchmark)

CRONOS-Benchmark is a controlled, synthetic benchmark for evaluating **counterfactual physical consistency** in video world models. It tests whether generative video models correctly simulate three fundamental physical event types — **object falling**, **object collision**, and **object occlusion** — across diverse scenes, objects, and viewpoints.

Each sequence provides ground-truth RGB frames, depth maps, and segmentation masks, along with a 5-frame conditioning clip and a `metadata.json` file, enabling standardised evaluation of any video generation model that accepts image or video conditioning.

## Dataset Details

### Dataset Description

CRONOS-Benchmark dataset contains **675 sequences** rendered in a controlled simulation environment. Each sequence captures one of three physical event types with foreground objects placed in a realistic indoor or outdoor scene. Ground-truth multi-modal annotations (RGB, depth, mask) are provided for all 90 frames of every sequence.

- **License:** Apache 2.0


## Dataset Structure

### Taxonomy

| Axis | Values |
|------|--------|
| **Event type** | `fall`, `collision`, `occlusion` |
| **Scene** | `apartment`, `house`, `garden`, `pool`, `kitchen` |
| **Object** | `Bottle`, `Can`, `TennisBall`, `ToyTruck`, `SoccerBall` |
| **Appearance** | `appearance_0`, `appearance_1`, `appearance_2` |
| **Viewpoint** | `close`, `front`, `side`, `top` (fall/collision); `side` only (occlusion) |

### Sequence counts

| Event type | Sequences |
|------------|-----------|
| fall       | 300       |
| collision  | 300       |
| occlusion  | 75        |
| **Total**  | **675**   |

### Directory layout

```
dataset/
└── {event}/                          # fall | collision | occlusion
    └── {scene}/                      # apartment | house | garden | pool | kitchen
        └── {object}/                 # Bottle | Can | TennisBall | ToyTruck | SoccerBall
            └── {appearance}/         # appearance_0 | appearance_1 | appearance_2
                └── {viewpoint}/      # close | front | side | top
                    ├── rgb/          # 90 × frame_XXXX.jpg  (1920×1080, colour)
                    ├── depth/        # 90 × frame_XXXX.jpg  (1920×1080, depth)
                    ├── mask/         # 90 × frame_XXXX.jpg  (1920×1080, binary)
                    ├── movies/
                    │   ├── complete.mp4        # full 90-frame sequence
                    │   └── conditioning_5.mp4  # first 5 frames (conditioning input)
                    └── metadata.json           # sequence metadata and text prompt
```

### Modalities

| Modality | Format | Resolution | Description |
|----------|--------|------------|-------------|
| RGB | JPEG | 1920 × 1080 | Colour render of the scene |
| Depth | JPEG | 1920 × 1080 | Per-frame depth map |
| Mask | JPEG | 1920 × 1080 | Per object segmentation mask |
| `complete.mp4` | MP4 | 1920 × 1080 | Full ground-truth sequence (90 frames) |
| `conditioning_5.mp4` | MP4 | 1920 × 1080 | First 5 frames, used as model conditioning input |
| `metadata.json` | JSON | — | Sequence identifiers (`event`, `scene`, `object`, `appearance`, `view`) and a natural-language `prompt` describing the expected physical event |

### Dataset statistics

| Stat | Value |
|------|-------|
| Total sequences | 675 |
| Frames per sequence | 90 |
| Total JPEG frames | 182,250 |
| Total MP4 files | 1,350 |
| Uncompressed size | ~18.8 GB |



## Uses

### Direct Use

CRONOS-Benchmark is designed to evaluate video generation / world models on physical event generation. Given the first frame or a conditioning clip as input, a model is asked to generate the remaining frames of the sequence. 


### Out-of-Scope Use

- Real-world scene understanding (data is fully synthetic).

## Dataset Creation


### Source Data

#### Data Collection and Processing

All sequences were rendered synthetically using UnrealEngine5. All 3D assets were bought from the [Unreal Engine asset store (FAB)](https://fab.com).

#### Personal and Sensitive Information

The dataset contains no personal or sensitive information. All content is synthetic.

## Bias, Risks, and Limitations

- **Synthetic domain gap:** Models trained or tuned on real video may underperform relative to their true capability due to the synthetic appearance of the scenes.
- **Limited scene and object set:** Our combinatorial design limits the number of variations we can generate per visual factor. Therefore, only five scenes, five object and four views are included.



## Citation


```bibtex
@misc{begiristain2026cronos,
      title={CRONOS: Benchmarking Counterfactual Physical Consistency in Video Models}, 
      author={Le{\'o}n Begiristain and Olaf D{\"u}nkel and Adam Kortylewski},
      year={2026},
      eprint={2605.23699},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2605.23699}, 
}
```