Hugging Face researchers have introduced CoinVE-200K, a large-scale, high-quality dataset designed to advance compositional instruction-guided video editing. Unlike existing datasets that focus on single editing operations, CoinVE-200K supports multiple editing intents within the same video, requiring models to jointly understand and execute several atomic edits.
The dataset comprises 1080p video-editing pairs of up to 201 frames, with each sample involving 2 to 5 atomic editing operations. Instructions target humans, objects, and backgrounds, covering edit types such as addition, removal, modification, and stylization. A carefully designed generation and filtering pipeline ensures instruction faithfulness, visual quality, temporal consistency, and compositional diversity.
Alongside the dataset, the team presents CoinVE-Bench, a benchmark for evaluating compositional-instruction video editing across diverse subjects, operation types, and instruction complexities. They also introduce CoinVE-Edit, a 22B-parameter 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 demonstrate that CoinVE-Edit achieves strong performance in instruction following, compositional editing accuracy, visual quality, and temporal consistency. The project page, code, dataset, model, and benchmark are publicly available.