TainU/IV-Edit

该存储库包含**IV-Edit（指令视觉编辑）基准**和用于**RePlan框架**的训练数据。数据集旨在解决基于指令的图像编辑中的**指令视觉复杂性（IV-Complexity）**挑战，其中复杂的指令与杂乱或模糊的视觉场景交互。现有数据集通常以突出对象和直接命令为特征，而IV-Edit强调杂乱场景和需要细粒度视觉理解、复杂推理和精确区域控制的指令。数据集分为三个部分：`test`（官方IV-Edit基准，约800个手动验证的指令-图像对，专注于多样化和复杂的场景）、`train`（用于微调RePlan VLM规划器的训练数据）和`dev`（用于模型开发和超参数调整的验证集）。这些样本来自开源数据集（如COCO、LISA和TextAtlas），并经过筛选以满足IV-Complexity标准。

This repository contains the **IV-Edit (Instruction-Visual Editing)** Benchmark and the training data used for the **RePlan** framework. The dataset is designed to address the challenge of **Instruction-Visual Complexity (IV-Complexity)** in instruction-based image editing, where intricate instructions interact with cluttered or ambiguous visual scenes. While existing datasets often feature salient objects and direct commands, IV-Edit emphasizes cluttered scenes and instructions that require fine-grained visual understanding, complex reasoning, and precise region-level control. The dataset is organized into three splits: `test` (the official IV-Edit Benchmark, consisting of ~800 manually verified instruction-image pairs focusing on diverse, complex scenes), `train` (the training data used to fine-tune the RePlan VLM planner), and `dev` (a validation set for model development and hyperparameter tuning). These samples are derived from open-source datasets (such as COCO, LISA, and TextAtlas) and filtered to meet IV-Complexity standards.