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
cs.CV
2026-08-18
Tencent open-sources CoinVE-200K, 200K video-editing pairs with 2-5 atomic edits per prompt, plus CoinVE-Edit, a 22B model beating Seedance 2.0 on scope accuracy by 1.74.
Instruction-based video editors handle one edit at a time. Ask for "replace the shirt, remove the hat, clip sunglasses to the collar, and turn the wall into an oil painting" in a single prompt, and models start dropping instructions, editing the wrong region, or leaking a local style change into a global filter. Training data is the bottleneck: existing datasets pair videos with short single-intent instructions, target one subject at a time, and carry no systematic annotation for combining add, remove, replace, and stylize operations.
A Tencent team fills the gap with CoinVE-200K: 200K source-edited video pairs at 1080p and 81-201 frames, each prompt bundling 2 to 5 atomic edits (2.55 on average), with an average editing quality score of 4.85 out of 5. Dataset, benchmark (CoinVE-Bench), and a 22B baseline model (CoinVE-Edit) are all released.
The dataset is manufactured, not hand-annotated, in three stages.
CoinVE-Edit pairs Qwen3-VL-8B-Instruct (understanding) with Wan2.1-T2V-14B (generation), 22B in total. Two lightweight heads sit on the MLLM output: a Mask Predictor that outputs a spatiotemporal mask per instruction, answering "where to edit", and a GateNet that classifies each instruction as local or global, answering "whether the region constraint should bind". Instead of biasing attention with hard masks, each cross-attention block is wrapped in Q-Blending: weight 1 inside the edit region, soft attenuation to 0.7 (beta = 0.3) outside for local edits, full-frame effect for global ones, with per-instruction outputs summed by weight. Hard masks disturb the pretrained backbone's feature distribution; this soft wrapper adds zero trainable parameters inside the DiT, and the ablation backs the choice.
Training runs in three stages: feature alignment on 2M image-editing pairs (EditScore > 8), single-instruction training on mixed image and video data, then compositional fine-tuning on CoinVE-200K alone. The Mask Predictor and GateNet are first trained offline to an IoU of 0.71 and 95% gating accuracy before the full model unfreezes end to end. The whole run used 32 H200 GPUs.
CoinVE-Bench uses 361 videos strictly disjoint from training, with 11 metrics split between an MLLM checklist and specialized evaluators:
| Model | SA | SPA | EP | MN | CP |
| CoinVE-Edit (22B, open) | 87.97 | 89.45 | 89.60 | 95.30 | 90.83 |
| Seedance 2.0 (closed) | 85.34 | 87.71 | 88.08 | 92.87 | 93.91 |
| Kling O3 (closed) | 86.91 | 80.93 | 89.06 | 93.91 | 84.51 |
| SAMA (open) | 75.58 | 73.35 | 79.63 | 88.14 | 90.08 |
CoinVE-Edit leads all three editing-accuracy metrics, with scope accuracy 1.74 points above the best competitor, Seedance 2.0. The cost sits in fidelity: aesthetic quality 4.13 versus 4.47 (Seedance) and 4.49 (Kling), with scale consistency and content preservation also below Seedance. On single-instruction OpenVE-Bench, judged by Gemini 2.5 Pro, its 3.41 overall is the best among open models but trails closed-source Runway at 3.51.
The ablation separates the two ideas: concatenating all instructions into one long prompt without masks or Q-Blending drops SA from 87.97 to 82.61, and swapping Q-Blending for a hard mask bias lands at 83.35 while degrading scale and motion naturalness.
Two qualitative cases stand out. Given "clip sunglasses to the man's collar", most MLLM-based editors put the sunglasses on his face, because that is where training data puts sunglasses; CoinVE-Edit localizes first and places them correctly. When removing a cowboy hat, the mirror reflection disappears too, which suggests the Mask Predictor captures secondary affected regions beyond the primary target.
Real editing requests are almost always compositional, and that is the gap between demo and production for this class of tools. No training data existed at this scale for multi-instruction editing, and closed models are not stable here either: the paper shows frontier models leaking a local oil-paint restyle into a global filter, and Seedance 2.0 conjuring red beans into an empty bowl. CoinVE-200K ships data, benchmark, and an open-weight 22B baseline together, giving follow-up work a comparable starting point. The recipe of mask supervision plus soft attention routing also transfers beyond video, to any generative task where multiple instructions must hit different regions.
The authors flag the boundary themselves: no reference-based editing, fine-grained motion editing, or complex camera control, and long or highly interactive scenes remain hard.
Reading closely raises more concerns. The ground truth is synthetic: edited videos come out of Wan2.2-Animate/VACE and image editors, so the dataset inherits their ceiling, and Gemini filtering removes visible failures but not systematic synthesis artifacts; training a Wan-family backbone on Wan-family targets could amplify shared biases. The benchmark is self-built and judged by Gemini 3.6 Flash, while the training set was filtered by Gemini 2.5 Pro, so one model family both curated the data and referees the comparison. Trust the direction, not the decimal. A mask IoU of 0.71 means nearly a third of localization error is absorbed by Q-Blending's soft attenuation, with no separate analysis of heavy occlusion. And no human evaluation appears anywhere in the paper.