CoinVE-200K: A Large-Scale High-Quality Dataset for Compositional Instruction-Guided Video Editing

Fuchen Long, Cong Wang, Zitao Gao, Wenhao Zhong, Yu Cheng, Xiaolu Hou
Yan Li, Xiao Cao, Xinlong Sun^†, Xi Chen^✉, Yu Liu

^† Project Leader   ^✉ Corresponding Author

Smart Creation Platform Department, Online Video BU, Tencent

 

---

🌍 Introduction

Instruction-guided video editing has witnessed rapid progress recently, driven by large-scale datasets and diffusion-based video generation models. However, existing open-source datasets (e.g., ReCo-Data, OpenVE-3M) primarily focus on single-instruction editing — applying one editing operation (e.g., replace, add, remove, or stylize) to a source video at a time. This limits the practical capability of trained models in real-world scenarios where users often issue multiple editing instructions simultaneously for a single video. To bridge this gap, we introduce CoinVE-200K, a large-scale, high-quality dataset for compositional instruction-guided video editing. Each sample in CoinVE-200K contains multiple instructions applied to the same source video, along with per-instruction region masks and a combined mask indicating all edited regions. The dataset is constructed through a meticulously designed data pipeline with rigorous quality filtering, ensuring diversity in instruction combinations, editing types, and video content.

Key features of CoinVE-200K:

• Compositional Instructions: Each sample contains 2~5 instructions covering different editing operations (Replace, Add, Remove, Background Change, etc.) on different regions (subject, object, background).
• Region-Aware Masks: Per-instruction masks indicate the spatial region of each edit, and a combined mask aggregates all edited regions for holistic supervision.
• Large Scale & High Quality: 200K+ video-edit pairs with ~1.18M video files, totaling ~2.8 TB, sourced from diverse open-source video collections.
• Rich Annotations: Each sample includes structured fields — instruction text, operation type, object type, and corresponding mask video paths.

Demonstration of compositional instruction-guided video editing cases from CoinVE-200K.

📊 Dataset Statistics

Overview

Metric	Value
Video source	Subset of OpenVid-1M (i.e., OpenVidHD)
Total editing samples	200,916
Total video files	~1.18M
Total size	~2.8 TB
Video resolution	1080P
Max edited frames	201
Instructions per sample	2~5 (avg. 2.55)

File Distribution

Type	Directory	Shards	Files	Size
Source video	src_videos/	74	194,450	~1.56 TB
Edited video	tgt_videos/	41	200,916	~873 GB
Combined mask	combined_masks/	8	223,773	~158 GB
Instruction mask	instruction_masks/	10	558,717	~189 GB

📁 Dataset Structure

Directory Layout

CoinVE-200K/
├── src_videos/
│   ├── src_video_000.tar
│   ├── src_video_001.tar
│   └── ...
├── tgt_videos/
│   ├── tgt_video_000.tar
│   └── ...
├── combined_masks/
│   ├── combined_masks_000.tar
│   └── ...
├── instruction_masks/
│   ├── instruction_masks_000.tar
│   └── ...
└── metadata_coinve200k.jsonl

Tar Archive Structure

Each tar archive contains video files with relative paths:

src_video_000.tar
├── src_video_000/
│   ├── UWPBxW-hVEY_3_28to136.mp4
│   ├── VRWPztEQZwQ_67_0to117.mp4
│   └── ...

instruction_masks_003.tar
├── instruction_masks_003/
│   ├── UWPBxW-hVEY_3_28to136_86c7745f/
│   │   ├── instr_mask_01.mp4
│   │   ├── instr_mask_02.mp4
│   │   └── instr_mask_03.mp4
│   └── ...

metadata_coinve200k.jsonl Format

Each line is a JSON object representing one editing sample:

{
  "source_video_path": "src_videos/src_video_067/UWPBxW-hVEY_3_28to136.mp4",
  "edited_video_path": "tgt_videos/tgt_video_039/UWPBxW-hVEY_3_28to136_86c7745f.mp4",
  "instruction": [
    "Replace the white styrofoam takeout container with a brown cardboard clamshell burger box.",
    "Add a large silver metal fork resting on top of the french fries in the right side of the container.",
    "Replace the outdoor concrete sidewalk background with a dark wooden table surface."
  ],
  "instruction_operation": ["Replace", "Add", "Replace"],
  "instruction_object": ["subject", "object", "background"],
  "instruction_mask_video_paths": [
    "instruction_masks/instruction_masks_003/UWPBxW-hVEY_3_28to136_86c7745f/instr_mask_01.mp4",
    "instruction_masks/instruction_masks_003/UWPBxW-hVEY_3_28to136_86c7745f/instr_mask_02.mp4",
    "instruction_masks/instruction_masks_003/UWPBxW-hVEY_3_28to136_86c7745f/instr_mask_03.mp4"
  ],
  "combined_mask_video_path": "combined_masks/combined_masks_004/UWPBxW-hVEY_3_28to136_86c7745f_combined_mask.mp4"
}

📥 Download

Full Dataset

# Download all files from HuggingFace
hf download FireCRT/CoinVE-200K --repo-type dataset --local-dir ./CoinVE-200K

Partial Download

You can download specific file types to save bandwidth:

# Download metadata only
hf download FireCRT/CoinVE-200K metadata_coinve200k.jsonl --repo-type dataset --local-dir ./CoinVE-200K

# Download specific src_videos shards
hf download FireCRT/CoinVE-200K \
  src_videos/src_video_000.tar src_videos/src_video_001.tar \
  --repo-type dataset --local-dir ./CoinVE-200K

🔧 Usage

Load Metadata

import json

with open("CoinVE-200K/metadata_coinve200k.jsonl", "r") as f:
    samples = [json.loads(line) for line in f]

print(f"Total samples: {len(samples)}")
print(f"First sample instructions: {samples[0]['instruction']}")

Extract Tar Archives

The dataset is distributed as .tar shards. Extract them before loading:

# Extract all source video shards
for tar in src_videos/*.tar; do tar -xf "$tar" -C src_videos/; done

# Extract all other types similarly
for tar in tgt_videos/*.tar; do tar -xf "$tar" -C tgt_videos/; done
for tar in combined_masks/*.tar; do tar -xf "$tar" -C combined_masks/; done
for tar in instruction_masks/*.tar; do tar -xf "$tar" -C instruction_masks/; done

After extraction the directory layout matches the relative paths in metadata_coinve200k.jsonl:

CoinVE-200K/
├── src_videos/src_video_067/UWPBxW-hVEY_3_28to136.mp4
├── tgt_videos/tgt_video_039/UWPBxW-hVEY_3_28to136_86c7745f.mp4
├── instruction_masks/instruction_masks_003/UWPBxW-hVEY_3_28to136_86c7745f/instr_mask_01.mp4
├── combined_masks/combined_masks_004/UWPBxW-hVEY_3_28to136_86c7745f_combined_mask.mp4
└── metadata_coinve200k.jsonl

📜 Citation

If you find CoinVE-200K useful for your research, please cite our work:

@article{coinve200k,
  title={CoinVE-200K: A Large-Scale High-Quality Dataset for Compositional Instruction-Guided Video Editing},
  author={Long, Fuchen and Wang, Cong and Gao, Zitao and Zhong, Wenhao and Cheng, Yu and Hou, Xiaolu and Li, Yan and Cao, Xiao and Sun, Xinlong and Chen, Xi and Liu, Yu},
  journal={arXiv preprint arXiv:2608.17566},
  year={2026}
}

✉️ Contact

For any questions, issues, or collaborations, please feel free to contact longfc.ustc@gmail.com.

💖 Acknowledgement

Our source videos are sourced from the OpenVid-1M dataset (specifically the OpenVidHD subset). Thanks to the contributors of this impactful project!