data-archetype/cc12m-1mp-plus-realistic-bucketed-1024

该数据集是一个自包含、重新标注的导出版本，源自原始数据集的训练分割。它包含573,052个图像-标题对，已按宽高比分桶处理，目标基础分辨率为1024px，并包含768px和512px的备用桶以适应较小源图像。图像以JPEG格式存储，未进行重新调整大小、裁剪或重新编码；非JPEG输入则使用数据集的JPEG导出设置进行转换。每个样本包括一个图像、一个重新标注的英文标题和一个小型JSON元数据文件。数据集经过清理，移除了缺失图像占位符、空白/均匀图像、精确重复的pHash组以及视觉审查的低质量图像（如严重压缩、像素化、上采样或退化的源图像）。数据集以TAR分片形式打包，适用于文本到图像训练，并支持WebDataset加载。

This dataset is a self-contained, recaptioned export of the train split from the original dataset. It contains 573,052 image-caption pairs packaged as aspect-ratio-bucketed TAR shards for text-to-image training. Images are assigned to buckets targeting a 1024px base resolution, with 768px and 512px fallback buckets for smaller source images. Source JPEGs are stored without resizing, cropping, or re-encoding; non-JPEG inputs are converted to JPEG using the datasets JPEG export settings. Each sample contains an image, one recaptioned English caption, and a small JSON metadata file. The dataset has been cleaned to remove missing-image placeholders, blank/uniform images, exact duplicate pHash groups, and visually reviewed low-quality images such as heavily compressed, pixelated, upsampled, or otherwise degraded sources. It is designed for research and training workflows involving image-text models.