A1111 SDXL prompt emphasis experiment

TLDR: Repeating words in your prompt will have more effect than "(((word)))" for SDXL models.

A1111 SDXL prompt emphasis experiment

Putting emphasis on prompt words in A1111 like (bacon:1.1), (((bacon))), etc. does not work the same in SDXL models as we were used to from SD1.5 models. In fact it seems to have very little effect. Here is a small experiment to support my findings.

Positive emphasis

Compare the result when just repeating the word "bacon" over and over to the result when trying the traditional (bacon:X). The latter seems to do very little, except for when you go really high like (bacon:3). However, this will start hurting image stability and it still portrays a suboptimal amount of bacon. (For copious amounts of bacon I refer you to the image where I prompted for the word "bacon" a whopping 16 times)

Negative emphasis

Similarly, trying to see less of a certain object by using e.g. (eggs:0.5) will also have little to no effect, as shown by the image where I negatively emphasized the word eggs. However, contrary to positive emphasis, I have not yet found a workaround for negative emphasis. One would think that flooding the negative prompt would work. The image where I put the word "eggs" in the negative prompt 12 times did not significantly decrease the amount of eggs compared to an empty negative prompt. More testing required.

Now I hope you're as hungry as I am...