README.md

---
pretty_name: RememBench
language:
- en
task_categories:
- text-to-video
- image-to-video
tags:
- arxiv:2610.02153
- benchmark
- video-generation
- long-horizon
- memory
size_categories:
- n<1K
configs:
- config_name: t2v
  data_files:
  - split: test
    path: data/t2v/test.jsonl
- config_name: i2v
  data_files:
  - split: test
    path: data/i2v/test.jsonl
---

# RememBench

A benchmark for long-horizon revisit consistency in autoregressive video generation, introduced in **MosaiChunk**.

[Project page](https://mosaichunk.github.io/) · [Paper](https://arxiv.org/abs/2610.02153) · [Code](https://github.com/mosaichunk/MosaiChunk) · [Checkpoints](https://huggingface.co/mosaichunk/MosaiChunk) · [Video Viewer](https://mosaichunk.github.io/viewer/video.html)

## Data

| Subset | Scenes | Inputs | Contents |
|---|---:|---:|---|
| `t2v` | 100 | 100 | Four-segment prompts, timing, and seeds |
| `i2v` | 150 | 750 | Initial-frame sources, scene prompts, camera trajectories, and seeds |

Both subsets use the `test` split. I2V contains 50 indoor and 100 outdoor scenes. All scenes have 90°, 180°, and 360° rotations; outdoor scenes also have rotation with translation.

DL3DV initial frames are retrieved using the provided preparation script and are not bundled here.

## Usage

```python
from datasets import load_dataset

t2v = load_dataset("mosaichunk/RememBench", "t2v", split="test")
i2v = load_dataset("mosaichunk/RememBench", "i2v", split="test")
```

Seeds are stored as strings; use `int(sample["seed"])` when generating videos.

[Download, image preparation, data formats, and evaluation](docs/usage.md)

CLIP and LPIPS measure departure–revisit consistency. Included T2V frame annotations describe the paper's rollouts; new rollouts require their own revisit annotations.

## License

The release license is pending. DL3DV images remain subject to the [DL3DV Terms of Use](https://github.com/DL3DV-10K/Dataset/blob/main/License.md).