CoinVE-200K - Tencent's open-source large-scale combinatorial instruction video editing dataset

What is CoinVE-200K?

CoinVE-200K is a large-scale combined instruction video editing dataset open-sourced by Tencent Video's Intelligent Creation Team. It contains 200,000 pairs of 1080p video editing samples, each containing 2-5 atomic editing instructions, including adding, removing, modifying, and stylizing, covering multiple categories such as people, objects, and backgrounds, with a maximum length of 201 frames. CoinVE-200K is accompanied by the open-source 22-parameter model CoinVE-Edit and the evaluation set CoinVE-Bench, filling the triple gaps in combined instruction video editing in terms of data, models, and evaluation standards.

Main functions of CoinVE-200K

• Massive amounts of combined editing dataIt provides 200,000 sets of 1080p combined instruction video editing sample pairs to support multi-intent joint training of the model.
• Multi-atom operation coverageIt covers 6 atomic editing operations, including adding, removing, replacing, and stylizing, and can flexibly combine complex editing tasks.
• Full-scenario target supportThe editing targets span multiple subjects, including people, objects, and backgrounds, and possess rich scene and semantic diversity.
• Editable models out of the boxIt comes with the open-source 22B parameter CoinVE-Edit model, enabling precise execution of multiple editing commands at once.
• System Evaluation BenchmarksWe constructed the CoinVE-Bench benchmark suite to systematically evaluate combinatorial editing performance using 11 fine-grained metrics.

Technical Principles of CoinVE-200K

• Data generation and filteringBased on a predefined editing classification system, Qwen3.6-27B is used to analyze the source video content and identify editable subjects. Multiple atomic operations are combined to generate composite instructions. Then, SAM3/SAM2 is used to generate editing region masks. HunyuanImage-3.0 and Nano Banana are used to synthesize editing keyframes. The videos are then diffused into complete videos using Wan2.1-Animate or VACE. Gemini is used for secondary quality filtering to select qualified samples based on dimensions such as instruction compliance, visual quality, and temporal consistency.
• Multimodal semantic understandingCoinVE-Edit uses Qwen3-VL-8B-Instruct as its multimodal understanding module. It simultaneously receives source video frame sequences and multiple text editing instructions, jointly analyzes the video content and complex editing intentions, and extracts high-level semantic representations for combined editing, enabling the model to globally understand the relationships and constraints between multiple instructions.
• Edit-perceptual feature transformationThe hidden state output by MLLM is mapped to two types of edit-aware tokens through the Feature Connector: one type is learnable tokens bound to each instruction, and the other type is visual tokens that retain the original video information; among them, the instruction-related tokens are injected into the DiT backbone network as conditional signals to guide the generation process to accurately respond to different types of editing needs.
• Region decoupling attention mechanismThe Mask Predictor predicts the spatiotemporal editing region mask for each instruction, and GateNet generates dynamic gating coefficients. In the cross-attention layer of DiT, the Q-Blending mechanism differentiates the contribution of each instruction token and visual token to the query based on the mask and gating weights, so as to achieve accurate binding of different editing intentions in the corresponding regions, while suppressing interference from irrelevant regions and maintaining temporal consistency.
• Dual-track complementary evaluation systemCoinVE-Bench employs an evaluation framework that combines an MLLM-based checklist with a dedicated evaluator: the former uses Gemini 3.6 Flash to evaluate the correctness of instruction compliance from three dimensions: editing accuracy, physical naturalness, and semantic preservation; the latter uses Aesthetic Predictor, DOVER++, VisualQuality-R1, and optical flow field to measure aesthetic quality, technical quality, overall quality, and temporal stability, covering a total of 4 core dimensions and 11 indicators.

Follow us on WeChat and reply with "open source",join inAI open source project discussion group

How to use CoinVE-200K

• Download datasetAfter installing the ModelScope CLI, execute the following commands: modelscope download Command to fetch the full amount or download only as needed. metadata_coinve200k.jsonl Metadata will be tested first.
• Unzip the split filesExtract the downloaded tar files to the specified locations. src_videos、tgt_videos、combined_masks and instruction_masks Corresponding directory.
• Parse metadata: Using Python to read metadata_coinve200k.jsonlObtain the source video path, the edited video path, and 2–5 combined editing instructions for each sample.
• Call mask informationExtract from metadata instruction_mask_video_paths and combined_mask_video_pathObtain the spatiotemporal editing region mask corresponding to each instruction.
• Conduct model trainingThe parsed video pairs, combined instructions, and masks are input into the DiT framework to supervise the training and fine-tuning of the combined instruction video editing model.
• Reproduce baseline resultsDirectly load the official open source code. CoinVE-Edit Model weights, in CoinVE-Bench The performance of combined editing was verified on the evaluation set.

CoinVE-200K's core advantages

• Leading in both scale and quality:have 200,000 sets of 1080p The high-definition video editing pairs, with a maximum of 201 frames, are the largest combined instruction video editing dataset to date.
• Complex Combinatorial EditingEach sample contains 2–5 atomic editing instructions, supporting flexible and complex combinations of operations such as adding, removing, modifying, and stylizing.
• Strict quality filtrationThrough MLLM semantic understanding and secondary quality filtering pipeline, we ensure instruction compliance, visual quality, temporal consistency and integrity of unedited areas.
• Full-chain open source ecosystemIt simultaneously provides the 22B parameter CoinVE-Edit model and the CoinVE-Bench evaluation set, forming a complete closed loop of data-model-evaluation.
• Accuracy surpasses closed-source modelsIt surpasses mainstream closed-source solutions such as Seedance 2.0 and Kling O3 in terms of accuracy of combined instruction editing (SA/SPA/EP).
• Region decoupling editing mechanismBy combining Mask-based Conditioning with Q-Blending cross-attention, multiple instructions can be precisely bound to corresponding spatiotemporal regions while retaining irrelevant content.

CoinVE-200K project address

• Project official website：https://coinve200k.github.io/
• GitHub repository：https://github.com/coinve200k/CoinVE-200K
• HuggingFace model library：https://huggingface.co/datasets/FireCRT/CoinVE-200K
• arXiv technical paper：https://arxiv.org/pdf/2608.17566

Comparison of CoinVE-200K with similar competing products

Application scenarios of CoinVE-200K

• Film and television post-productionThe director uses natural language commands to complete multiple element replacements, background modifications, and style transfers in one go, significantly improving post-production efficiency.
• Advertising content iterationBrands can quickly perform batch editing on the same video footage, such as changing people's clothes, replacing products, and adjusting backgrounds, reducing the cost of reshooting.
• Short video creationCreators can input multiple editing intentions and achieve complex effects such as character removal, object addition, and image stylization with a single click, lowering the professional threshold.
• AI model trainingAs high-quality supervised data, it is used to train a large-scale video editing model with combined instructions, thereby improving the model's ability to understand and accurately execute multiple intentions.
• Evaluation benchmark research: Provides the academic community with a standardized CoinVE-Bench to systematically evaluate the performance of video editing models in terms of instruction compliance, physical naturalness, and visual quality.