README.md

---
license: cc-by-nc-4.0
task_categories:
- text-to-video
- video-to-video
language:
- en
tags:
- video-editing
- instruction-based-editing
- video
size_categories:
- 1M<n<10M
---

<div align="center">

# Goku: A Million-Scale Universal Dataset and Benchmark for Instruction-Based Video Editing

[![arXiv](https://img.shields.io/badge/arXiv-2606.30599-b31b1b.svg?logo=arxiv&logoColor=white)](https://arxiv.org/abs/2606.30599)
[![Project Page](https://img.shields.io/badge/Project-Page-1f8ceb.svg?logo=googlechrome&logoColor=white)](https://flying-sky999.github.io/Goku.github.io/)
[![License](https://img.shields.io/badge/License-CC%20BY--NC%204.0-4caf50.svg)](https://creativecommons.org/licenses/by-nc/4.0/)

</div>

GOKU-2M is a large-scale, unified **instruction-based video-editing** dataset covering **10 editing tasks**. Each sample provides a source video, an edited target video, and one or more natural-language instructions describing the edit.

## 📦 Repositories

> ⚠️ Because a single Hugging Face account has a free storage quota of about **8.7&nbsp;TB**, the dataset is split across **two repositories**:

| 🗂️ Repository | 💾 Size | 🎬 Tasks |
|---|---|---|
| 🔵 [`bigfacing/GOKU-2M`](https://huggingface.co/datasets/bigfacing/GOKU-2M) | **5.11&nbsp;TB** | add, remove, swap, alter, reference-based add, reference-based swap, camera motion, style transfer |
| 🟢 [`Goku-2M/GOKU-2M`](https://huggingface.co/datasets/Goku-2M/GOKU-2M) | **4.54&nbsp;TB** | subject movement, multi-step composite editing |

## Tasks

<div align="center">
<img src="assets/teaser.png" width="100%" alt="GOKU-2M teaser: dataset distribution and per-task editing examples">
</div>

| Folder | Task | Description |
|---|---|---|
| `add` | Add | Add a new object into the scene |
| `remove` | Remove | Remove an object from the scene |
| `swap_alter` | Swap / Alter | Replace an object, or alter its attributes |
| `reference_add` | Reference Add | Add an object specified by a reference image |
| `reference_swap` | Reference Swap | Replace an object with one from a reference image |
| `camera` | Camera Motion | Apply a camera movement (pan / tilt / zoom / arc / translate) |
| `style_transfer` | Style Transfer | Restyle the whole video |
| `subject_movement` | Subject Movement | Edit the motion / action of the subject |
| `multi_task` | Multi-step Composite | Chained edits, e.g. subject edit followed by camera motion |

## Repository Layout

Each task is a folder containing sharded `.tar` files:

```
<task>/
├── <task>.videos.part001.tar
├── <task>.videos.part002.tar
├── ...
└── <task>.jsons.tar
```

Concatenate and extract all shards of a task. After extraction you get two folders:

```
videos/
└── <case_id>/
    ├── source.mp4          # the input video
    ├── edited.mp4          # the edited result (single-step tasks)
    ├── reference.jpg       # only for reference_add / reference_swap
    └── ...                 # step videos for multi_task (i2v.mp4, camera.mp4, ...)
jsons/
└── combine_json/
    └── <case_id>_all.json  # annotation for the case
```

`<case_id>` is shared between the `videos/` and `jsons/` folders — the JSON's video paths are relative to the task root (e.g. `videos/<case_id>/source.mp4`).

## Annotation Format

Every case has one JSON at `jsons/combine_json/<case_id>_all.json`. There are two schemas.

### 1. Single-step tasks (add, remove, swap_alter, reference_add, reference_swap, style_transfer, subject_movement)

Flat schema:

```json
{
  "case_id": "6jqn3hpk8elm4zd5_..._refadd",
  "step": "reference_add",
  "source_video": "videos/6jqn3hpk8elm4zd5_..._refadd/source.mp4",
  "edited_video": "videos/6jqn3hpk8elm4zd5_..._refadd/edited.mp4",
  "reference_image": "videos/6jqn3hpk8elm4zd5_..._refadd/reference.jpg",
  "instruction_en": "Add a white, modern ceramic vase ... The object to add is shown in the reference image.",
  "resolution": { "width": 1280, "height": 720 },
  "fps": 25
}
```

Common fields: `case_id`, `step`, `source_video`, `edited_video`, `instruction_en`, `resolution`, `fps`.
Task-specific fields:
- `reference_image` — present for `reference_add` and `reference_swap`.
- `long_instruction_en` — a more detailed instruction, present for `swap_alter` and `reference_swap`.
- `instruction_zh` / `long_instruction_zh` — Chinese instructions, present for some tasks (e.g. `swap_alter`, `subject_movement`).
- `source_caption` / `edited_caption` — full-scene captions of the input and result, present for `subject_movement`.

### 2. Multi-step tasks (camera, multi_task)

Composite schema with an ordered `steps` list and per-step `pairs`:

```json
{
  "case_id": "1f23d8486d3547e8_translate_down",
  "combo_name": "subject_camera",
  "steps": ["i2v", "camera"],
  "total_instruction_en": "Make the trees lean further inward ..., then move the camera downward",
  "total_instruction_zh": "...，然后镜头向下移动",
  "source_video": "videos/1f23d8486d3547e8_translate_down/source.mp4",
  "final_video": "videos/1f23d8486d3547e8_translate_down/camera.mp4",
  "pairs": [
    { "step": "i2v",    "source_video": "...", "edited_video": "...", "instruction_en": "...", "...": "..." },
    { "step": "camera", "source_video": "...", "edited_video": "...", "instruction_en": "...", "cam_type": 8, "cam_name": "..." }
  ],
  "resolution": { "width": 736, "height": 704 },
  "fps": 25
}
```

Key fields: `steps` (ordered edit types), `total_instruction_en/zh` (overall instruction), `source_video` → `final_video` (start / end of the chain), and `pairs` (each intermediate `source→edited` step with its own instruction; camera steps additionally carry `cam_type` / `cam_name`).

## Usage

Download a single task and extract it:

```bash
# Install the HF CLI
pip install -U "huggingface_hub[cli]"

# Download one task folder (e.g. "remove" from repo 1)
hf download bigfacing/GOKU-2M --repo-type dataset \
    --include "remove/*" --local-dir ./GOKU-2M

# Extract all shards of that task
cd GOKU-2M/remove
for f in *.tar; do tar -xf "$f"; done
# -> produces videos/ and jsons/
```

Load and iterate over annotations:

```python
import json, glob, os

task_root = "GOKU-2M/remove"          # folder containing videos/ and jsons/
for jp in glob.glob(os.path.join(task_root, "jsons/combine_json/*.json")):
    ann = json.load(open(jp))
    src = os.path.join(task_root, ann["source_video"])
    if "pairs" in ann:                # multi-step task
        dst = os.path.join(task_root, ann["final_video"])
        instruction = ann["total_instruction_en"]
    else:                             # single-step task
        dst = os.path.join(task_root, ann["edited_video"])
        instruction = ann["instruction_en"]
    ref = ann.get("reference_image")  # only reference_add / reference_swap
    # src -> dst under `instruction` (+ optional reference image)
```

## Notes

- Released under **CC-BY-NC-4.0**: free for non-commercial research use with attribution.
- Videos contain no audio.
- File and case names are non-sensitive identifiers; only video content and JSON annotations carry semantic information.
- For the full per-task details, browse the two repositories linked above.

## Links

- 📄 Paper: https://arxiv.org/abs/2606.30599
- 🌐 Project Page: https://flying-sky999.github.io/Goku.github.io/

## Citation

If you find GOKU-2M useful for your research, please cite:

```bibtex
@article{liang2026goku,
  title={Goku: A Million-Scale Universal Dataset and Benchmark for Instruction-Based Video Editing},
  author={Liang, Sen and Wang, Cong and Yu, Zhentao and Guan, Fengbin and Zhou, Zhengguang and Hu, Teng and Zhang, Youliang and Zhou, Yuan and Li, Xin and Lu, Qinglin and others},
  journal={arXiv preprint arXiv:2606.30599},
  year={2026}
}
```