Stable Diffusion 2: The Good, The Bad and The Ugly

Generally, Stable Diffusion 1 is trained on LAION-2B (en), subsets of laion-high-resolution and laion-improved-aesthetics. The datasets is then filtered for explicit pornographic material, using the LAION-NSFW classifier with punsafe 0.1 and an aesthetic score 4.5. In other words, Stable Diffusion 2 used a larger, NSFW-filtered datasets for training. Also, the second phase of training is based on images with resolution higher or equal to 512x512. The new release ensures that StabilityAI resolves the long-standing legal problems related to child pornography and deep fakes. Also, the new model generates better images in certain areas.