donghao-zhou/ThinkV2V-150K

📌 Dataset Summary

ThinkV2V-150K is a complex-prompt video editing dataset for instruction-guided video editing. It is filtered from OpenVE-HQ-1M and further processed through prompt rewriting, converting direct editing instructions into more detailed, reasoning-oriented video editing requests.

This release does not duplicate video files. The videos are hosted by OpenVE-3M. ThinkV2V-150K provides CSV metadata with rewritten prompts and OpenVE-3M-compatible relative video paths.

🗂️ File Structure

The released dataset is organized as follows:

ThinkV2V-150K/
├── README.md
└── csv_files/
    ├── background_change.csv
    ├── global_style.csv
    ├── local_add.csv
    ├── local_change.csv
    └── local_remove.csv

• csv_files/: complex-prompt metadata for the five released editing categories.
• Video files are not included in this repository. Please download them from OpenVE-3M.

🔗 Using With OpenVE-3M

ThinkV2V-150K is intended to be used together with the extracted OpenVE-3M videos. The released training metadata is restricted to samples that can be reliably aligned to public OpenVE-3M video paths. A typical local layout is:

datasets/
├── OpenVE-3M/
│   ├── videos/
│   │   ├── background_change/
│   │   ├── global_style/
│   │   ├── local_add/
│   │   ├── local_change/
│   │   └── local_remove/
│   └── csv_files/
│       └── ...
└── ThinkV2V-150K/
    ├── README.md
    └── csv_files/
        ├── background_change.csv
        ├── global_style.csv
        ├── local_add.csv
        ├── local_change.csv
        └── local_remove.csv

The video and original_video fields use the same relative-path style as OpenVE-3M:

background_change/a130108912c20457607fdcb9d880005d.mp4
global_style/b888304ee7030294ff23806c605ca138.mp4
local_change/6c7b7857b88bd5470532c7b4a2da2c4c.mp4

The training dataloader should read metadata from ThinkV2V-150K and resolve the video and original_video fields using the configured OpenVE-3M video path.

🧾 Data Fields

Each row in the category CSV files contains one video editing triplet:

Field	Type	Description
video	path	Relative path to the edited video under the extracted OpenVE-3M videos/ directory.
prompt	string	Rewritten complex editing instruction.
original_video	path	Relative path to the source video under the extracted OpenVE-3M videos/ directory.

The CSV files follow the OpenVE-3M metadata format and use semicolon delimiters:

video;prompt;original_video

📊 Data Size

Category	Samples
background_change	24,261
global_style	39,665
local_add	36,214
local_change	25,732
local_remove	30,653
Total	156,525

🔍 Example Sample

An example entry from csv_files/background_change.csv is shown below:

{
  "video": "background_change/a130108912c20457607fdcb9d880005d.mp4",
  "prompt": "Swap the current pool-side backdrop for a lively open woodland clearing, where tiny glowing winged insects flicker gently, overhead foliage rustles softly in a light breeze, and shifting sunbeams filter through the treetops to create moving dappled patches across the mossy forest floor. Layer in soft, distant calls of winged forest dwellers to enhance the calm, serene atmosphere, while leaving the gray-haired man and the large white arctic bear completely stationary in their positions.",
  "original_video": "background_change/38fbe37b56ee0ac595ffe994d7639e4c.mp4"
}

The paths above refer to files from the extracted OpenVE-3M video archive, not files stored in this ThinkV2V-150K repository.

🎯 Intended Uses

ThinkV2V-150K is intended for research and training of instruction-guided video editing systems under complex prompts. Typical use cases include training or fine-tuning video editing models with reasoning-oriented editing instructions.

This release is intended for research and non-commercial use.

💻 How to Load

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

from pathlib import Path
import pandas as pd

openve3m_video_root = Path("datasets/OpenVE-3M/videos")
thinkv2v_root = Path("datasets/ThinkV2V-150K")

meta = pd.read_csv(thinkv2v_root / "csv_files" / "background_change.csv", sep=";")

sample = meta.iloc[0].to_dict()
edited_video = openve3m_video_root / sample["video"]
source_video = openve3m_video_root / sample["original_video"]

print(sample["prompt"])
print(edited_video)
print(source_video)

⚖️ License and Ethics

ThinkV2V-150K is derived from OpenVE-HQ-1M and OpenVE-3M and is released under CC BY-NC 4.0 for research and non-commercial use.

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