LAION-5B
An extremely large-scale image-text pair dataset released by the German non-profit organization LAION, containing approximately 5.85 billion image-text pairs. It serves as a core training foundation for mainstream image generation models such as Stable Diffusion.
Timeline (last 90 days)
Mainstream image-text paired datasets (such as LAION-5B) have already reached the billions scale, whereas audio-natural language description paired datasets of equivalent scale have yet to appear.
Large-scale image-text datasets (such as LAION-5B and Common Crawl-derived data) are primarily collected from publicly available internet resources and exhibit uneven distribution issues, including English content dominance and overrepresentation of Western aesthetic standards.
LAION-5B数据集包含约50亿张图像-文本对，主要来源于互联网网页图片的alt标签
Stable Diffusion was trained on the LAION-5B dataset, which contains approximately 5.8 billion image-text pairs and was the largest publicly available multimodal dataset at the time.
Stable Diffusion基于LAION-5B数据集训练，采用U-Net架构作为去噪骨干网络，配合CLIP文本编码器实现文本到图像的语义映射
All Facts (5)
Stable Diffusion was trained on the LAION-5B dataset, which contains approximately 5.8 billion image-text pairs and was the largest publicly available multimodal dataset at the time.
60%UnverifiedMainstream image-text paired datasets (such as LAION-5B) have already reached the billions scale, whereas audio-natural language description paired datasets of equivalent scale have yet to appear.
50%UnverifiedLarge-scale image-text datasets (such as LAION-5B and Common Crawl-derived data) are primarily collected from publicly available internet resources and exhibit uneven distribution issues, including English content dominance and overrepresentation of Western aesthetic standards.
50%UnverifiedLAION-5B数据集包含约50亿张图像-文本对，主要来源于互联网网页图片的alt标签
50%UnverifiedStable Diffusion基于LAION-5B数据集训练，采用U-Net架构作为去噪骨干网络，配合CLIP文本编码器实现文本到图像的语义映射
50%