I just found out that there is an article area today, so I thought about it and wrote something casually. These are all techniques that I couldn't think of myself, so please don't laugh at me.
First of all, my data set never removes the background
In themselves, they can increase the fit with the characters, and the most important point is that if your ps skills are insufficient, it is likely that the ai will learn things that it should not learn.
In addition, occasionally there are some pictures with a gradient background, but since I use a cheap monitor, I didn't find any traces left on the picture when I processed the watermarke,Such as the color in the picture below,please watch carefully
Hateful thing is that AI often learns this weird stuff faster than you want it to learn
At this time, rather than smearing it out, I am more used to selecting some objects that AI can recognize for covering, such as the white frame mosaic used by some other creators
However, this thing is usually not recognized well, you need to add tags manually, and this kind of tag is sometimes like an easter egg, maybe some unlucky guy will trigger it accidentally, who knows?Most importantly, if you use too much of an item, even if you don't use the tag, it may appear with the picture
At this time, I will use some objects that are easier to be recognized by AI and will not make the picture too obtrusive to cover the watermark
In short, in my lora, I have added fish...chibi cat...flowers...band-aid, etc. The most important point is that it matches the content of the picture and does not repeat it many times in your data set
This method can also remove many objects that you don't want AI to learn, but appear too many times and have complex patterns.
So,write something casually for a while, and see you later.
OH，yes,maybe you can ask,me,the cover's large watermark can remove use what method?
Just turn off the layers.