I repaired the eyes, but it still as example generates green or red eyes when prompted for brown eyes. However, it no longer produces damaged eyes in my tests. Additionally, it now creates fewer blurry, white, or messy genitals. Unfortunately, I had to remove a bit of the SDXL merge to make these repairs. This part was responsible for making the backgrounds nice and providing knowledge outside of my normal anime dataset and tags. As a result, the backgrounds are a bit worse and less detailed, and the model seems a little less stable due to my merging back to a version without eye problems and my additional training on it.
Using training for better backgrounds resulted in much simpler and worse backgrounds, so I dropped that attempt and created this compromise model, which still generates good backgrounds. You can still use version 1.3, but if you don't care much about backgrounds and don't mind slightly less colorful images, you can use this model instead.
What is better:
Eye artifacts should be now very rare
Fewer incorrect eye colors.
Better formed out pupils (including special pupils like hearts-shaped pupils)
Reduced blurry, white, or messy genitals.
What is worse:
Colors are more muted.
Worse and less detailed backgrounds and landscapes.
Minimal more issues with hands and composition.
Some sfw concepts are a little bit less precise
I am now working on my natural language model to generate NSFW captions for images and on SD3.
I repaired the eyes, but it still as example generates green or red eyes when prompted for brown eyes. However, it no longer produces damaged eyes in my tests. Additionally, it now creates fewer blurry, white, or messy genitals. Unfortunately, I had to remove a bit of the SDXL merge to make these repairs. This part was responsible for making the backgrounds nice and providing knowledge outside of my normal anime dataset and tags. As a result, the backgrounds are a bit worse and less detailed, and the model seems a little less stable due to my merging back to a version without eye problems and my additional training on it.
Using training for better backgrounds resulted in much simpler and worse backgrounds, so I dropped that attempt and created this compromise model, which still generates good backgrounds. You can still use version 1.3, but if you don't care much about backgrounds and don't mind slightly less colorful images, you can use this model instead.
What is better:
Eye artifacts should be now very rare
Fewer incorrect eye colors.
Better formed out pupils (including special pupils like hearts-shaped pupils)
Reduced blurry, white, or messy genitals.
What is worse:
Colors are more muted.
Worse and less detailed backgrounds and landscapes.
Minimal more issues with hands and composition.
Some sfw concepts are a little bit less precise
I am now working on my natural language model to generate NSFW captions for images and on SD3.