Overview
LAION-Aesthetics is a curated subset of LAION-5B where images are scored and filtered for aesthetic quality. This filtering proved crucial for training Stable Diffusion, demonstrating that quality-selected subsets outperform raw scale for generative models.
What’s In It
The dataset contains approximately 600 million image-text pairs from LAION-5B that scored above a threshold on a trained aesthetics predictor. Higher-threshold subsets provide increasingly curated selections of visually appealing images.
How It’s Used
LAION-Aesthetics was the key training set for Stable Diffusion with the aesthetic filtering proving essential for generating visually appealing outputs. It demonstrated that curating for quality is more important than raw scale for image generation.
Controversies
The aesthetics predictor inherits biases about what constitutes beautiful images, potentially favoring certain cultural aesthetics. The same CSAM and copyright concerns affecting LAION-5B apply to this subset.