[Sample] V-CAPE: Visual Consistency Assessment for Product Image Editing

V-CAPE Data Sample A small stratified sample from the V-CAPE benchmark, intended for anonymous reviewer quality inspection. The sample contains 3 examples per class (label × rejection reason) from each of the two dataset splits: Split Description Classes Samples vcape-r Real e-commerce editing pairs accepted, Geometry Artifacts, Wrong Orientation, Irrelevant Objects, Texture/Lighting/Color Issues, Background Issues, Generated-Main Image Mismatch 21 vcape-s Synthetic pairs rendered from 3D assets accepted, wrong_orientation, wrong_product_different_pt, wrong_product_same_pt 12 Files File Description data-00000-of-00001.arrow HuggingFace dataset in Arrow format containing all 33 samples dataset_info.json HuggingFace dataset metadata (features, num rows) state.json HuggingFace dataset state file croissant-sample.json Croissant metadata describing the dataset schema image_downloader.py Standalone script to download vcape-r source images by physical_id visualize_datasample.py Generates a self-contained HTML report for visual inspection report.html Pre-generated HTML visualization with the data sample (open in any browser) Dataset Schema Each row contains the following columns: Column Type Description split string Origin split: `vcape-r` or `vcape-s` physical_id string Source image identifier for vcape-r rows (empty for vcape-s) xsource_image Image Embedded source image for vcape-s rows (null for vcape-r) xtarget_image Image Generated/rendered target image object_description string Natural-language description of the product product_type string Product category (e.g., "sofa") pose string Target orientation / pose label input_prompt string Editing prompt for the transformation label string Human quality label: `accepted` / `rejected` (vcape-r) or `accept` / `reject` (vcape-s) rejection_reason string Structured rejection reason from a predefined taxonomy Quick Start 1. Install dependencies ```bash pip install datasets Pillow requests ``` 2. Load the dataset ```python from datasets import load_from_disk ds = load_from_disk(".") # run from this directory print(ds) print(ds[0]) ``` 3. Download vcape-r source images ```bash # Download all vcape-r source images into source_images/ python image_downloader.py # Or download specific images python image_downloader.py --pids 91FmmLIE3GL 71tLzgZoBQL ``` 4. Visualize the sample Either open the pre-generated `report.html` in a browser, or regenerate it: ```bash python visualize_datasample.py ``` This produces a self-contained HTML file with image thumbnails, metadata, and filter controls for split, label, and rejection reason.# V-CAPE 数据样本（V-CAPE Data Sample） 本数据集为V-CAPE基准测试集的分层小样本子集，专供匿名审稿人开展质量核验使用。 该样本集从两个数据集划分中，按「标签×驳回原因」的类别各抽取3个样本： | 数据集划分 | 描述 | 类别 | 样本数 | | --- | --- | --- | --- | | vcape-r | 真实电商编辑样本对 | accepted（合规）、Geometry Artifacts（几何异常）、Wrong Orientation（朝向错误）、Irrelevant Objects（无关物体）、Texture/Lighting/Color Issues（纹理/光照/色彩问题）、Background Issues（背景问题）、Generated-Main Image Mismatch（生成主图不匹配） | 21 | | vcape-s | 基于3D资产渲染的合成样本对 | accepted（合规）、wrong_orientation（朝向错误）、wrong_product_different_pt（产品错误且视角不同）、wrong_product_same_pt（产品错误但视角相同） | 12 | ## 文件说明 | 文件 | 描述 | | --- | --- | | data-00000-of-00001.arrow | 采用Arrow格式存储的HuggingFace数据集，包含全部33个样本 | | dataset_info.json | HuggingFace数据集元数据文件（包含特征信息、样本行数等） | | state.json | HuggingFace数据集状态文件 | | croissant-sample.json | 描述数据集 schema 的Croissant元数据文件 | | image_downloader.py | 用于通过physical_id下载vcape-r划分源图像的独立脚本 | | visualize_datasample.py | 生成可独立运行的HTML可视化报告以辅助质量核验的脚本 | | report.html | 预生成的数据集样本可视化HTML报告，可直接在任意浏览器中打开 | ## 数据集结构 每条数据样本包含以下字段： | 字段名 | 数据类型 | 说明 | | --- | --- | --- | | split | 字符串 | 数据所属的数据集划分：`vcape-r` 或 `vcape-s` | | physical_id | 字符串 | vcape-r划分的源图像标识符（vcape-s划分此字段为空） | | xsource_image | 图像（Image） | vcape-s划分的嵌入式源图像（vcape-r划分此字段为null） | | xtarget_image | 图像（Image） | 生成/渲染得到的目标图像 | | object_description | 字符串 | 产品的自然语言描述 | | product_type | 字符串 | 产品类别（例如："沙发"） | | pose | 字符串 | 目标朝向/姿态标签 | | input_prompt | 字符串 | 用于图像变换的编辑提示词 | | label | 字符串 | 人工质量标注：vcape-r划分采用 `accepted` / `rejected`，vcape-s划分采用 `accept` / `reject` | | rejection_reason | 字符串 | 来自预定义分类体系的结构化驳回原因 | ## 快速开始 1. 安装依赖环境 bash pip install datasets Pillow requests 2. 加载数据集 python from datasets import load_from_disk ds = load_from_disk(".") # 在当前目录下执行该代码以加载数据集 print(ds) print(ds[0]) 3. 下载vcape-r划分的源图像 bash # 将所有vcape-r源图像下载至source_images/目录下 python image_downloader.py # 或仅下载指定标识符对应的图像 python image_downloader.py --pids 91FmmLIE3GL 71tLzgZoBQL 4. 可视化数据集样本 您可以直接在浏览器中打开预生成的`report.html`文件，或重新生成该可视化报告： bash python visualize_datasample.py 该脚本将生成一份包含图像缩略图、元数据以及针对划分、标签和驳回原因的筛选控件的独立HTML报告。