NOT COMPATIBLE WITH V1 BASED MODELS.

i did this bc i was bored and i find writing ridiculous negative prompts funny and nobody had done it yet with wd 1.4. to say i was shocked at the results is an understatement. it can turn extremely simple prompts like "1girl" into masterpieces without having to actually say "masterpiece" or anything other than "by wd14neg" in negatives. so you can all stop complaining that wd 1.4 isn't as good as anything v3 now. i never rly share my stuff bc i hate attention but i just had to this time.

trained on 500+ images generated with wd 1.4 anime from an extremely exaggerated negative prompt in positives (seriously it was like 2000 tokens) and quality enhancers in negatives.

wd14neg is the best and can stand on its own, but may improve with wd14neg-2 and/or wd14neg-3. combining all of them may result in deepfrying. to solve this i would either lower your cfg or deemphasize it.

the embedding may struggle with eliminating extra limbs and fingers since i mainly focused on blurriness, flat colors, and jpeg artifacts to fix embeddings of mine with crappy datasets i don't have the executive function to redo but i plan to make another embedding focusing on those.