The quality and diversity of instruction-based video editing datasets are steadily improving, yet existing datasets mainly focus on single editing operations and fall short in supporting compositional instruction-guided video editing. In particular, multiple editing intents must be jointly understood and faithfully executed within the same video. To address this issue, we introduce CoinVE-200K, a large-scale, high-quality dataset for Compositional Instruction-Guided Video Editing. CoinVE-200K contains 1080p video-editing pairs of up to 201 frames, covering diverse compositional scenarios where each sample involves 2 to 5 atomic editing operations. The instructions target humans, objects, and backgrounds, and cover edit types such as addition, removal, modification, and stylization. All samples are built through a carefully designed generation and filtering pipeline to ensure instruction faithfulness, visual quality, temporal consistency, and compositional diversity. We also introduce CoinVE-Bench, a benchmark for compositional-instruction video editing across diverse subjects, operation types, and instruction complexities. Furthermore, we present CoinVE-Edit, a 22B compositional video editing model built upon Wan2.1-T2V-14B and Qwen3-VL-8B-Instruct. CoinVE-Edit disentangles region-aware attention for different editing instructions, enabling precise multi-region editing while preserving irrelevant content and temporal coherence. Experiments on CoinVE-Bench show that CoinVE-Edit achieves strong performance in instruction following, compositional editing accuracy, visual quality, and temporal consistency.
Figures & tables
Figure 1: Demonstration of compositional instruction-guided video editing cases from our proposed CoinVE-200K.
Figure 2: Data construction pipeline of CoinVE-200K.
Figure 3: Data statistics of CoinVE-200K.
Figure 4: An overview of the proposed CoinVE-Edit framework.
Dimension
Metric (Abbr.)
Range
Evaluator
Description
MLLM-Based (Gemini 3.6 Flash)
Editing Accuracy (per Instruction)
Semantic Accuracy (SA)
[0,100]
MLLM + checklist
Correct execution of the intended edit semantics
Scope Accuracy (SPA)
[0,100]
MLLM + checklist
Correct edit localization without leakage or interference
Editing Persistence (EP)
[0,100]
MLLM + checklist
Temporal consistency of the edit throughout the video
Physical Naturalness
Appearance Naturalness (AN)
[0,100]
MLLM + checklist
Natural blending of lighting, shadows, textures, and style
Scale Consistency (SC)
[0,100]
MLLM + checklist
Plausibility of the edited object’s scale and perspective
Table 1: Evaluation matrix of CoinVE-Bench. The proposed metrics are grouped by evaluator type into MLLM-based checklist evaluation and specialized evaluator-based evaluation with task-specific models, covering four evaluation dimensions and eleven fine-grained metrics.
Model
Overall
Global Style
Background Change
Local Change
Local Remove
Local Add
Subtitle Edit
Runway
3.51
3.72
2.62
4.18
4.16
2.78
3.62
VACE
1.55
1.49
1.55
2.07
1.46
1.26
1.48
Ditto
2.25
4.01
1.68
2.03
1.53
1.41
2.81
OpenVE-Edit
2.57
3.16
2.36
2.98
1.85
2.15
2.91
VINO
2.91
4.05
1.67
3.12
3.32
2.66
2.64
OmniWeaving
3.01
3.94
2.03
3.49
2.93
2.06
3.58
Table 2: Single-instruction video editing comparison on OpenVE-Bench evaluated by Gemini 2.5 Pro.
Figure 5: Quality comparison of single-instruction video editing on OpenVE-Bench.
Model
Edit. Acc.
Phys. Natural.
Seman. Pres.
Video Quality
SA
SPA
EP
AN
SC
MN
CP
AQ
TQ
CQ
TS
Seedance 2.0
85.34
87.71
88.08
93.19
95.84
92.87
93.91
4.47
19.55
4.41
0.62
Kling O3
86.91
80.93
89.06
92.55
90.30
93.91
84.51
4.49
18.36
4.37
0.61
VACE
3.98
17.15
6.50
26.69
13.82
15.21
87.83
4.05
17.59
4.11
0.62
Ditto
34.69
36.41
40.85
35.96
47.79
38.48
51.98
3.59
17.24
3.96
0.67
VINO
83.63
66.75
89.06
78.09
82.34
85.91
61.70
4.06
17.28
4.08
0.68
Table 3: Compositional-Instruction Video Editing Comparisons on CoinVE-Bench.
Figure 6: Quality comparison of compositional-instruction video editing on CoinVE-Bench.
Variant
Components
Edit. Acc.
Phys. Natural.
Seman. Pres.
Mask
Q-Blending
SA
SPA
EP
AN
SC
MN
CP
Instr.-Concat.
×
×
82.61
84.43
89.17
86.52
90.02
92.17
89.30
Q-Bias
✓
×
83.35
85.49
88.19
87.08
89.64
91.95
90.13
CoinVE-Edit
✓
✓
87.97
89.45
89.60
91.85
91.17
95.30
90.83
Table 4: Performance comparison among different variants of CoinVE-Edit on CoinVE-Bench.
Figure 7: Visualization of predicted editing mask of each instruction for two editing examples.
While recent text-guided video editing models excel at elementary tasks (e.g., style transfer, object insertion), real-world user requests are highly compositional. A single prompt often demands multiple coupled edits, such as modifying subjects, actions, and camera views, while strictly preserving unrelated spatiotemporal content. Existing benchmarks, heavily constrained by isolated edits and coarse global metrics, fail to diagnose how models handle such complex workflows. To address this gap, we introduce CoVEBench, a compositional video editing benchmark comprising 416 curated source videos, 626 multi-point editing instructions, and 9,990 fine-grained checklist items. Covering diverse editing dimensions, CoVEBench evaluates models via MLLM-judged instruction compliance and video fidelity, alongside automated metrics for video quality. Extensive experiments reveal that compositional editing remains a profound challenge: current models frequently omit edits, violate preservation constraints, or introduce artifacts when handling multiple operations simultaneously. CoVEBench provides a challenging, diagnostic testbed to advance video editing toward realistic user workflows.
Jiangtao Wu, Jiaming Wang, Yiwen He +7
NJU-LINK Team, Nanjing University · Kling Team, Kuaishou Technology
While instruction-based video editing has seen significant progress, joint audio-visual editing remains constrained by the absence of dedicated datasets and benchmarks. To bridge this gap, we present JAVEdit-100k, the first large-scale, high-quality dataset tailored for instruction-guided joint audio-visual editing. Focusing on human-centric videos, JAVEdit-100k comprises approximately 100K editing triplets spanning five distinct categories, including subject editing and speech editing. This dataset is rigorously constructed via four meticulously designed generation pipelines, seamlessly paired with an agent-in-the-loop quality control mechanism. Furthermore, to address the lack of standardized evaluation within the field, we introduce JAVEditBench, a comprehensive benchmark featuring curated source videos and human-aligned instructions across all editing categories. Finally, we propose JAVEdit, a pioneering baseline model for instruction-guided joint audio-visual editing. Experiments show that \model\ outperforms all baselines on five of six evaluation metrics.
Yinan Chen, Chuming Lin, Zhennan Chen +12
1Zhejiang University · 2Tencent Youtu Lab · 3Nanjing University +3
Instruction-based video editing has witnessed rapid progress, yet current methods often struggle with precise visual control, as natural language is inherently limited in describing complex visual nuances. Although reference-guided editing offers a robust solution, its potential is currently bottlenecked by the scarcity of high-quality paired training data. To bridge this gap, we introduce a scalable data generation pipeline that transforms existing video editing pairs into high-fidelity training quadruplets, leveraging image generative models to create synthesized reference scaffolds. Using this pipeline, we construct RefVIE, a large-scale dataset tailored for instruction-reference-following tasks, and establish RefVIE-Bench for comprehensive evaluation. Furthermore, we propose a unified editing architecture, Kiwi-Edit, that synergizes learnable queries and latent visual features for reference semantic guidance. Our model achieves significant gains in instruction following and reference fidelity via a progressive multi-stage training curriculum. Extensive experiments demonstrate that our data and architecture establish a new state-of-the-art in controllable video editing. All datasets, models, and code is released at https://github.com/showlab/Kiwi-Edit.