Rapidata/bananamark-dataset

Bananamark是一个用于评估图像生成模型在香蕉主题生成任务上表现的基准数据集。该数据集包含2161个训练样本，每个样本包含提示词、两幅由不同模型生成的香蕉图像、模型名称、人类评分权重和详细结果。通过Rapidata平台收集了超过20,000条人类偏好数据，评估标准是选择哪张图像的香蕉更好。测试了包括FLUX 2、Seedream 4、Google Nano Banana等在内的11个主流图像生成模型，并提供了完整的性能排行榜。结果表明FLUX 2 Flex表现最佳，而尽管命名为Nano Banana的Google模型仅排名第三和第四。

Bananamark is a benchmark dataset for evaluating the performance of image generation models on banana-themed generation tasks. The dataset contains 2,161 training examples, each consisting of a prompt, two generated banana images (from different models), model names, human preference weighted scores, and detailed results. Over 20,000 human preferences were collected via the Rapidata platform, with the evaluation criterion being which image has better bananas. The benchmark tested 11 major image generation models including FLUX 2, Seedream 4, Google Nano Banana, etc., and provides a comprehensive leaderboard. Results show FLUX 2 Flex performed best, while Googles Nano Banana (despite its name) only ranked 3rd and 4th.