---
license: cc-by-nc-4.0
task_categories:
- text-to-video
- video-to-video
language:
- en
tags:
- video-editing
- instruction-guided-video-editing
- reasoning
- mllm
- benchmark
size_categories:
- n<1K
---

## 📌 Benchmark Summary

**ThinkV2V-Bench** is a complex instruction benchmark for **instruction-guided video editing**. It is derived from [OpenVE-Bench](https://huggingface.co/datasets/Lewandofski/OpenVE-Bench), while focusing on editing categories where multimodal large language model (MLLM) thinking is more important for understanding and executing the requested transformation.

ThinkV2V-Bench contains **308** samples from five spatial editing categories:

```text
background_change    59
global_style         58
local_add            67
local_change         65
local_remove         59
```

The retained categories are selected because they require stronger visual grounding, object-level understanding, spatial reasoning, and instruction interpretation, making them better suited for evaluating the role of MLLM thinking in video editing.

## 🗂️ File Structure

The released benchmark is organized as follows:

```text
ThinkV2V-Bench/
├── README.md
├── benchmark_videos.csv
└── videos/
    ├── 0000_global_style_Apply_the_Impression.mp4
    ├── 0001_global_style_Apply_the_Ukiyoe_ani.mp4
    └── ...
```

- `benchmark_videos.csv`: entry file that indexes all benchmark samples.
- `videos/`: source videos used for video-to-video editing.

## 🧾 Data Fields

Each row in `benchmark_videos.csv` contains the following fields:

| Field | Type | Description |
| :--- | :---: | :--- |
| `edited_type` | string | Editing category for the sample. |
| `prompt` | string | Complex editing instruction rewritten from the original OpenVE-Bench prompt. |
| `original_video` | path | Relative path to the source video. |

The CSV field names are aligned with the OpenVE-Bench metadata format:

```text
edited_type,prompt,original_video
```

## 🔍 Example Sample

An example entry from `benchmark_videos.csv` is shown below:

```json
{
  "edited_type": "global_style",
  "prompt": "Give the video the aesthetic of the 19th-century art movement that focused on capturing the fleeting effects of light with small, visible brushstrokes.",
  "original_video": "ThinkV2V-Bench/videos/0000_global_style_Apply_the_Impression.mp4"
}
```

## 🎯 Intended Uses

ThinkV2V-Bench is intended for **benchmarking and analysis** of instruction-guided video editing systems under complex prompts. Typical use cases include:

- Evaluating how well a model follows complex video editing instructions.
- Studying whether MLLM-based thinking improves video editing under indirect or reasoning-oriented prompts.
- Analyzing failure cases in prompt interpretation, spatial grounding, and temporal consistency.

This release is primarily intended for **research and benchmarking**, rather than unrestricted training use.

## 🧩 Benchmark Tasks

ThinkV2V-Bench focuses on video-to-video editing:

```text
source video + complex editing instruction -> edited video
```

The benchmark covers five editing categories:

- **Global Style**: transform the whole video into a target visual style.
- **Background Change**: replace or transform the background while preserving foreground content.
- **Local Change**: replace or modify specified local objects or attributes.
- **Local Remove**: remove specified objects from the video.
- **Local Add**: add specified objects or elements into the video.

## 📏 Evaluation Protocol

Each benchmark sample provides a source video and a complex editing prompt. A model should generate an edited result for each row in `benchmark_videos.csv` while preserving the row order or otherwise recording the corresponding source video path.

For compatibility with OpenVE-style evaluation pipelines, model outputs can be organized in a CSV by adding an edited-result column to the metadata, for example:

```text
edited_type,prompt,original_video,edited_result_path
```

The `edited_result_path` column is the default model-output column expected by the inference and evaluation scripts. It is added after running inference and is not part of the released benchmark metadata.

The evaluation setup should compare the source video and generated edited video under the provided complex prompt.

## 💻 How to Load

You can load the benchmark metadata with standard Python tools such as `pandas`.

```python
from pathlib import Path
import pandas as pd

root = Path("ThinkV2V-Bench")
meta = pd.read_csv(root / "benchmark_videos.csv")

print(meta.columns.tolist())
sample = meta.iloc[0].to_dict()
print(sample)

video_path = root / sample["original_video"].replace("ThinkV2V-Bench/", "", 1)
print(video_path)
```

## ⚖️ License and Ethics

ThinkV2V-Bench is derived from OpenVE-Bench and is released under **CC BY-NC 4.0** for research and non-commercial benchmarking use.

Users should apply the benchmark responsibly and avoid using it for harmful, deceptive, or privacy-violating applications. Generated editing results are determined by model behavior and user-provided instructions.