The style of Zlixon.Trigger word is <code>zlix0n</code>. If you get black-and-white images, consider putting <code>monochrome</code> in the negatives. <code>sketch</code> doesn't seem to have much effect, but you may try it too.I'd also like to take this opportunity to rant a bit. CivitAI's latest update to the on-site trainer has made it so that it refuses to process the dataset if it doesn't like some tags in it, but it doesn't actually tell you what those tags are. First of all, I'm against this kind of censorship altogether (accurate tagging is equally useful for invoking and preventing concepts), but on a more personal level, this just makes the dataset preparation significantly more tedious. Now I have to go through the gelbooru tags, make guesses at which ones offend the censorship, remove them and see if anything's changed. Not only does it degrade the model's accuracy (I assume the effect is negligible), but it takes my valuable time. If anyone from CivitAI is reading this, I'd like to propose that you at least give an option to automatically remove tags you're unhappy with, or, even better, just disable dataset tag censorship, considering you're already filtering all posted and generated images.If you like the style, please support the <a target="_blank" rel="ugc" href="https://www.pixiv.net/en/users/71120789">original author</a>.

Files with SHA256: b8629e3351951d7e2b6008eef361621496a814aed204670704d6ad1619041ff6