CoinVE-Bench

   

---

✨ Highlights

• Compositional Editing Benchmark: CoinVE-Bench is a benchmark dedicated to compositional (multi-instruction) video editing evaluation, featuring 361 test cases with 2–5 instructions per sample.
• Region-Aware Evaluation: Per-instruction masks enable fine-grained checks of whether each edit is confined to the correct region.
• Multi-Dimensional Metrics: Comprehensive evaluation across editing accuracy, physical naturalness, and semantic preservation.
• VLLM-Based Scoring: Leveraging vision-language models (e.g., Gemini) for human-aligned assessment.

📊 Benchmark Statistics

Metric	Value
Total test samples	361
Instructions per sample	2–5
Checklist questions	4,131
Evaluation dimensions	3 (Editing Accuracy / Physical Naturalness / Semantic Preservation)

Scoring Dimensions & Metrics

All metrics are scored by Gemini 3.6 Flash (default; configurable via --model) using a checklist-based VLLM protocol.

Dimension	Metric (Abbr.)	Range	# Questions	Description
Editing Accuracy (per instruction)	Semantic Accuracy (SA)	[0, 100]	1,081	Correct execution of the intended edit semantics.
	Scope Accuracy (SPA)	[0, 100]	379	Correct edit localization without leakage or interference.
	Editing Persistence (EP)	[0, 100]	923	Temporal consistency of the edit throughout the video.
Physical Naturalness	Appearance Naturalness (AN)	[0, 100]	356	Natural blending of lighting, shadows, textures, and style.
	Scale Consistency (SC)	[0, 100]	521	Plausibility of the edited object's scale and perspective.
	Motion Naturalness (MN)	[0, 100]	447	Plausibility of motion and physical interactions.
Semantic Preservation	Content Preservation (CP)	[0, 100]	424	Preservation of non-edited regions, objects, and structures.

Scoring formulas:

• ACC-type metrics (SA / SPA / EP / AN / SC / MN): score = num_correct / num_questions × 100
• Score-type metric (CP): score = mean(score) × 10 (each question is scored 0–10, then scaled to [0, 100])
• Dimension score = arithmetic mean of its metric score_100 values

Question Types

Type	# Questions	Videos Uploaded	Answer Format
Single True/False (Single-TF)	2,524	Edited Video only	Yes / No
A/B Multiple Choice Question (AB-MCQ)	609	Edited Video only	A / B / A and B
Dual True/False (Dual-TF)	574	Source Video + Edited Video	Yes / No
Score Multiple Choice Question (Score-MCQ)	424	Source Video + Edited Video	0–10 score

📁 Dataset Structure

CoinVE-Bench/
├── src_videos/
│   ├── 9CicJDFN1TA_9_0to183.mp4
│   ├── gWG0LiuZnFQ_8_0to137.mp4
│   ├── JwVDrRmyoXc_51_0to172.mp4
│   └── ...
└── checklist_json/
    └── coinve-bench-361-checklist.json

Source videos are named <youtube_id>_<clip_index>_<start>to<end>.mp4 and are referenced by the src_video field of each case in the checklist (e.g. src_videos/9CicJDFN1TA_9_0to183.mp4).

🔧 Evaluation

Evaluation is a two-stage pipeline: (1) run your editing model on the 361 source videos to produce edited videos, (2) score them with the Gemini-based evaluator.

1. Run Inference on CoinVE-Bench

Generate one edited video per test case using your model. For CoinVE-Edit, use the inference scripts in the CoinVE-Edit model repository.

2. Prepare Edited Videos

The evaluator expects one flat directory of edited MP4 files with a strict naming convention:

edited_videos/your_model/
├── 0.mp4
├── 1.mp4
├── 2.mp4
└── ...
└── 360.mp4

• Format: MP4 (.mp4).
• Naming: <case_id>.mp4 where <case_id> is the integer id field of each case in checklist_json/coinve-bench-361-checklist.json (range 0–360, no zero-padding).
• Source video pairing: Video A (original) is read from case["src_video"] inside the checklist; you do not need to place source videos in this directory.
• Missing videos: if a case's edited video is absent, all its questions are recorded with model_answer = null and counted as failures in the final score.

Note for CoinVE-Edit users: infer_coinve_bench.py saves a standalone edited video at tgt_videos/<case_id>.mp4 per case, already in the {id}.mp4 naming convention expected by the evaluator. You can pass the tgt_videos/ directory directly as --gen-video-dir.

3. Run Gemini Evaluation

The evaluation code is provided in the CoinVE-Edit repository. Set your Gemini API key and run:

export GEMINI_API_KEY=your_key_here
python eval_coinbench_gemini_public.py \
    --input ./checklist_json/coinve-bench-361-checklist.json \
    --gen-video-dir ./edited_videos/your_model \
    --prompt-dir ./system_prompts \
    --out-dir ./results/your_model \
    --workers 8 \
    --model gemini-3.6-flash \
    --timestamp

The script performs everything end-to-end — Gemini calls, per-question scoring, dimension/metric aggregation, and report writing. No separate aggregation step is needed.

📈 Performance Comparisons

Metric groups: SA / SPA / EP → Edit. Acc. · AN / SC / MN → Phys. Natural. · CP → Seman. Pres.

Model	SA	SPA	EP	AN	SC	MN	CP
Seedance 2.0	85.34	87.71	88.08	93.19	95.84	92.87	93.91
Kling O3	86.91	80.93	89.06	92.55	90.30	93.91	84.51
VACE	3.98	17.15	6.50	26.69	13.82	15.21	87.83
Ditto	34.69	36.41	40.85	35.96	47.79	38.48	51.98
VINO	83.63	66.75	89.06	78.09	82.34	85.91	61.70
OmniWeaving	59.67	55.94	61.11	54.49	66.03	65.10	75.09
KiWiEdit	76.50	69.92	80.28	78.37	78.50	80.76	70.31
SAMA	75.58	73.35	79.63	83.43	83.88	88.14	90.08
CoinVE-Edit	87.97	89.45	89.60	91.85	91.17	95.30	90.83

📜 Citation

If you find CoinVE-Bench 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 [email protected].

💖 Acknowledgement

Our benchmark construction and evaluation protocol are inspired by ReCo-Bench and OpenVE-Bench. Thanks to the contributors of all these remarkable projects!