Composite Being

Using the default negative prompt (DMP) in conjunction with the basic author prompt has resulted in some glorious stuff! It’s interesting because I had read in the stable diffusion subreddit that the type of terms in the DMP aren’t really effective, instead suggesting that the idea of a universal DMP is mistaken.

The theory went that because the images scraped into the training datasets originate from websites, they end up with a TON of junk search engine optimization (SEO) terms captioned to them that are totally unrelated to the actual picture.

The solution then was to write your prompt, set to maximum negative influence, see what you get, and then describe the results and use that as the neg prompt. Example: you want a noir detective cat. Do a run with prompt set to -3 or whatever your max neg val is. Result is pizza and chicken wings. Now run original prompt positive with pizza and chicken wings as neg prompt instead of DMP.

Yet our DMP indubitably improves things greatly!