README.md

---
license: apache-2.0
pretty_name: AVISet
language:
  - en
tags:
  - video-editing
  - mask-guided-editing
  - video-captioning
  - instruction-following
  - multimodal
size_categories:
  - 10K<n<100K
---

# AVISet

AVISet is a large-scale dataset for mask-guided, text-conditioned video editing. Each sample pairs a source video with a temporally aligned mask video that identifies the editable subject or region. Natural-language captions describe the source content, and the test split additionally provides an editing prompt describing the desired edited result.

AVISet is the official dataset released with **[Audio-sync Video Instance Editing with Granularity-Aware Mask Refiner](https://arxiv.org/abs/2512.10571)**.

The dataset contains 73,505 samples across training, validation, and test splits. All media is packaged into independently extractable TAR shards for reliable downloading and large-scale data loading.

## AVI-Edit

- **Paper:** [arXiv:2512.10571](https://arxiv.org/abs/2512.10571)
- **Code:** [github.com/suimuc/AVI-Edit-Framework](https://github.com/suimuc/AVI-Edit-Framework)
- **Project page:** [hjzheng.net/projects/AVI-Edit](https://hjzheng.net/projects/AVI-Edit/)

## Dataset Summary

| Split | Samples | Source videos | Mask videos | TAR shards | TAR size |
| --- | ---: | ---: | ---: | ---: | ---: |
| Training | 71,505 | 71,505 | 71,505 | 36 | 170.67 GB |
| Validation | 1,000 | 1,000 | 1,000 | 1 | 2.38 GB |
| Test | 1,000 | 1,000 | 1,000 | 1 | 2.30 GB |
| **Total** | **73,505** | **73,505** | **73,505** | **38** | **175.35 GB** |

The training and validation splits provide source captions. The test split also contains `editing_prompt`, which describes the intended transformation of the masked subject or region.

## Repository Structure

```text
AVISet/
├── README.md
├── training.csv
├── validating.csv
├── testing.csv
├── training/
│   ├── part_01.tar
│   ├── part_02.tar
│   ├── ...
│   └── part_36.tar
├── validating/
│   └── part_01.tar
└── testing/
    └── part_01.tar
```

Each sample contains two MP4 files with matching numeric identifiers:

```text
000000.mp4
000000_mask.mp4
000001.mp4
000001_mask.mp4
...
```

- `<id>.mp4` is the source video.
- `<id>_mask.mp4` is its temporally aligned mask video.
- White mask pixels identify the editable foreground or subject.
- Black mask pixels identify regions intended to remain unchanged.

The source and mask videos have matching frame counts, frame rates, durations, and spatial dimensions. Source videos may contain an audio track, while mask videos contain video only. Media properties such as resolution and frame rate can vary across samples.

## CSV Schema

### Training and validation

`training.csv` and `validating.csv` contain the following fields:

| Field | Type | Description |
| --- | --- | --- |
| `path` | string | Path to the source video relative to the extracted dataset root, for example `training/000000.mp4`. |
| `mask_path` | string | Path to the corresponding mask video, for example `training/000000_mask.mp4`. |
| `caption` | string | Natural-language description of the source video's subjects, actions, appearance, and scene. |

### Test

`testing.csv` contains one additional field:

| Field | Type | Description |
| --- | --- | --- |
| `path` | string | Path to the source video relative to the extracted dataset root, for example `testing/000000.mp4`. |
| `mask_path` | string | Path to the corresponding mask video, for example `testing/000000_mask.mp4`. |
| `caption` | string | Natural-language description of the original source video. |
| `editing_prompt` | string | Text description of the desired edited video, especially the intended transformation of the masked subject or region. |

An abbreviated test record looks like:

```csv
path,mask_path,caption,editing_prompt
testing/000000.mp4,testing/000000_mask.mp4,"A young man appears to be engaged in a conversation...","A young woman with long dark hair appears to be engaged in a conversation..."
```

## Download

Download the complete dataset with the Hugging Face CLI:

```bash
huggingface-cli download suimu/AVISet \
  --repo-type dataset \
  --local-dir AVISet
```

To download only selected files or splits, use `--include`. For example:

```bash
# Validation metadata and media only
huggingface-cli download suimu/AVISet \
  --repo-type dataset \
  --include "validating.csv" "validating/*" \
  --local-dir AVISet

# Test metadata and media only
huggingface-cli download suimu/AVISet \
  --repo-type dataset \
  --include "testing.csv" "testing/*" \
  --local-dir AVISet
```

## Extraction

Extract each split into a directory with the same name as the CSV path prefix:

```bash
mkdir -p data/training data/validating data/testing

for shard in AVISet/training/part_*.tar; do
  tar -xf "$shard" -C data/training
done

for shard in AVISet/validating/part_*.tar; do
  tar -xf "$shard" -C data/validating
done

for shard in AVISet/testing/part_*.tar; do
  tar -xf "$shard" -C data/testing
done

cp AVISet/training.csv AVISet/validating.csv AVISet/testing.csv data/
```

The resulting layout is:

```text
data/
├── training.csv
├── validating.csv
├── testing.csv
├── training/
│   ├── 000000.mp4
│   ├── 000000_mask.mp4
│   └── ...
├── validating/
│   ├── 000000.mp4
│   ├── 000000_mask.mp4
│   └── ...
└── testing/
    ├── 000000.mp4
    ├── 000000_mask.mp4
    └── ...
```

Each TAR shard can be extracted independently. Files in different shards use unique identifiers within their split, so all shards for a split can be extracted into the same directory.

## Citation

If you find AVISet or AVI-Edit useful for your research, please cite:

```bibtex
@article{avi-edit,
  title={Audio-sync Video Instance Editing with Granularity-Aware Mask Refiner},
  author={Zheng, Haojie and Weng, Shuchen and Liu, Jingqi and Yang, Siqi and Shi, Boxin and Wang, Xinlong},
  journal={arXiv preprint arXiv:2512.10571},
  year={2025}
}
```

## License

The repository declares the Apache License 2.0. Users are responsible for verifying that their intended use also complies with any rights and restrictions applicable to the underlying media.