>>101608148Like most things, dataset and tagging changes have been the most effective. For example, the red thing around her neck, which I tagged as "red collar" (I might try a different tag because sometimes it tries to morph into a dog collar). It's prevalent in the dataset, so it would show up even when not prompted or in the negative. So I took the dataset images that didn't have the collar, tagged them with "no collar", and added an extra repeat for those images. It's not perfect, but it works more consistently than not. I also did something similar for the gloves/gauntlets. I tagged the images without them with "bare hands".
In picrel, you can see that without using the "no collar" prompt, the shape of the red collar is still there. But prompting "no collar" gets the intended look.
In the future, I'll probably try masked training, but that seems like a lot of effort.
