terminusresearch/ideogram-25k

该数据集主要包含从Ideogram服务中提取的约27,000张图像，用于微调或训练文本到图像模型和分类器，以及分析Ideogram用户偏见。数据集的创建者使用自定义的Selenium应用程序监控Ideogram服务并保存数据，并通过SHA256哈希进行去重。数据集的描述由Microsoft Florence2生成，但存在来自单一合成源（Llava 34B captioner）的偏见。

This dataset contains approximately 27,000 images pulled from Ideogram, primarily used for fine-tuning or training text-to-image models and analyzing Ideogram user bias. The filenames are SHA256 hashes of the image data, used for integrity verification. The dataset was created to obtain high-quality typography data, with a synthetic data source free of copyright concerns. Data collection and processing were conducted using a custom Selenium application, monitoring the Ideogram service and saving posts to disk, with deduplication by SHA256 hash. The datasets bias stems from a single synthetic source, the Llava 34B captioner.