B-Free
收藏资源简介:
B-Free数据集由那不勒斯腓特烈二世大学和Google DeepMind合作创建,旨在解决AI生成图像检测中的偏差问题。该数据集包含51,000张真实图像和309,000张合成图像,总计360,000条数据。真实图像来源于COCO数据集,合成图像通过Stable Diffusion 2.1模型生成,确保了语义对齐。数据集的创建过程中,采用了自条件重建和内容增强技术,以避免语义和编码格式上的偏差。B-Free数据集主要应用于AI生成图像的检测和验证,旨在提高模型在未见过的生成模型上的泛化能力和鲁棒性。
B-Free dataset was co-developed by the University of Naples Federico II and Google DeepMind, aiming to mitigate biases in AI-generated image detection. This dataset consists of 51,000 real images and 309,000 synthetic images, with a total of 360,000 data samples. The real images are sourced from the COCO dataset, while the synthetic images are generated via the Stable Diffusion 2.1 model, with semantic alignment guaranteed. During the creation of this dataset, self-conditioned reconstruction and content augmentation techniques were employed to avoid biases in semantics and encoding formats. The B-Free dataset is mainly utilized for AI-generated image detection and validation, with the objective of improving the generalization capability and robustness of models across unseen generative models.




