Unsolved Issues & Advice There are still problems with the portrayal of details such as eyes, hands, and feet, which is largely due to the fact that the LAION dataset used by Stable Diffusion contains a lot of inappropriate data, such as Tom and Jerry, who are not five-fingered creatures. In the three models I released this time, I tried to eliminate some of the negative effects brought by the "laion-aesthetics v2 5+" used by Stable Diffusion 1.4. Can you believe that in the "laion-aesthetics v2 5+" data, if you search for Apple iPod, you will get many meme images of apples that can be eaten with headphones inserted, with a high aesthetics score. This is obviously wrong; on the other hand, Stable Diffusion 1.2 is a better version. If you want to continue training a Stable Diffusion model with anime style, please avoid Stable Diffusion 1.4+ version as possible. When you see through its releasing history, you will find that the subsequent versions mostly involve struggles and interests among the founding teams, and there is no good quality assurance. It is not that the higher the version, the better. My personal opinion is that Stable Diffusion 1.2 is the best version. Certains Certains Certains