CoinVE-200K: A Large-Scale High-Quality Dataset for Compositional Instruction-Guided Video Editing

The Problem: Video Editing Gets Complicated

Current datasets for instruction-based video editing focus on single, isolated edits—add a hat, remove a person, change the background color. Real editing workflows are messier. A user might say "make the person taller, give them a red jacket, and add snow falling in the background" all at once. Existing datasets don't prepare models for this compositional challenge, where multiple edits must happen simultaneously in the same video while staying faithful to the instructions and keeping everything temporally coherent.

The authors present CoinVE-200K, a dataset of 201,000 video-editing pairs designed explicitly for compositional instruction-guided editing. Each sample involves 2–5 atomic editing operations stacked together. Videos are high-resolution (1080p), up to 201 frames long, and cover edits to humans, objects, and backgrounds across addition, removal, modification, and stylization categories.

How the Dataset Was Built

The authors used a "carefully designed generation and filtering pipeline" to ensure quality—though the abstract doesn't specify the mechanics. The key is that they filtered for four criteria: instruction faithfulness (the edited video actually matches what was requested), visual quality (the output looks realistic), temporal consistency (no flickering or discontinuities across frames), and compositional diversity (varied combinations of operations). This careful curation distinguishes CoinVE-200K from simpler synthetic datasets.

Alongside the dataset, the authors introduce CoinVE-Bench, a benchmark for evaluating compositional editing across different subjects, operation types, and instruction complexities.

The Model and Results

The authors built CoinVE-Edit, a 22-billion-parameter model combining a text-to-video backbone (Wan2.1-T2V-14B) with a vision-language encoder (Qwen3-VL-8B-Instruct). The novel architectural contribution is "region-aware attention disentanglement"—essentially, the model learns to focus different attention heads on different spatial regions and editing tasks, so that modifying a person's jacket doesn't accidentally distort the background or break temporal continuity.

On CoinVE-Bench, CoinVE-Edit shows strong performance across instruction following, compositional editing accuracy, visual quality, and temporal consistency. Specific metrics aren't listed in the abstract, so the relative improvement margin is unclear.

Why This Matters

Compositional editing is the practical frontier. Users don't think in isolated operations; they think in complex, layered intentions. By creating a large-scale dataset that enforces multiple simultaneous edits with strict quality gates, the authors establish a new benchmark for the field. The model's region-aware attention mechanism is a concrete architectural answer to a real problem: how to edit multiple regions independently without interference. This work bridges the gap between single-operation datasets and the multi-step editing scenarios that real users need.