FractalAIResearch/ProductConsistency-Benchmark
收藏资源简介:
ProductConsistency Benchmark 是一个高质量的数据集,用于基于指令的、以产品为中心的英文图像编辑。每个记录是一个评估元组:包括一个输入产品图像和一个单一的编辑指令,描述广告风格的场景变化。数据集包含870个样本,由174个合成产品图像(覆盖8个产品类别)组成,每个图像有5个不同的编辑指令。图像强调清晰可读的包装文本和一致的品牌标识;指令编写时注重在现实营销风格编辑下保持身份一致性。基准图像在类别和渲染文本长度范围内均匀覆盖,使用与更广泛的ProductConsistency发布相同的合成产品生成和基于OCR的文本验证流程。对于每个选定的产品图像,语言模型生成五个针对该图像的独特编辑指令。为确保数据集质量,最终的人工验证步骤确认每个基准图像中的产品文本与预期真实情况匹配,然后再配对和发布指令。
The ProductConsistency Benchmark is a high-quality dataset for instruction-based, product-centric image editing in English. Every record is one evaluation tuple: an input product image plus a single edit instruction describing an advertisement-style scene change. The dataset contains 870 samples formed from 174 synthetic product images (8 product categories) with 5 distinct edit instructions per image. Images emphasize legible on-pack text and consistent branding; instructions are written to stress identity preservation under realistic marketing-style edits. Benchmark images are drawn with uniform coverage over categories and over the range of rendered text lengths, using the same synthetic product generation and OCR-based text verification pipeline as the broader ProductConsistency release. For each selected product image, a language model generates five edit instructions that are unique to that image. To ensure dataset quality, a final human verification step confirms that on-product text in each benchmark image matches the intended ground truth before instructions are paired and released.




