furproxy/furry_dataset_captions_img_dump

该数据集主要包含以兽迷主题为主的图像及其详细标注。大部分标注（至少占总数的70%）经过人工审核和编辑，基于Gemini 3 Flash生成的标注，并通过多裁剪处理和组合的流程完成。图像来源于e621、civit和booru，以md5命名存储。目前数据集包含约7.5k条标注，遵循一种通用格式，包括场景介绍、角色细节、互动以及背景元素。作者还提到在空间描述（如“看向肩膀后方”/“背后”/“前面”）方面偶尔存在小问题。

This dataset primarily consists of furry-themed images with detailed captions. Most of the captions (at least 70% of the total) are human-reviewed and edited, based on Gemini 3 Flash captions, involving a pipeline with multi-crop passes and combination. The images are sourced from e621, civit, and booru, stored with md5 names. The dataset currently contains around 7.5k captions, following a general format that includes scene introduction, character details, interactions, and background elements. The author also notes occasional minor issues with spatial descriptions (e.g., looking over shoulder/behind/in front of).