Weighting Hack for #stablediffusion I've been experimenting with punctuation and the more brackets you enclose a word with, the less weight it seems to have. Prompt: "close up face female portrait, Vampire," with parentheses around vampire. Anyone tried this?

Aug 7, 2022 · 4:13 PM UTC

14
15
7
165
As I already had the notebook open, here's an animation towards full vampire and back. I still need to test this more, but it does seem to be something that could be quite useful.
2
2
29
0
Replying to @TomLikesRobots @delta_sauce
Interesting, yet we can put in the same prompt and it gives out variations every single time, so is there an element of that happening here too or do you think it's because of the added prompt device and how that affects interpretation?
1
2
This was using the same seed and the only thing that was changed was the brackets in the prompt. The 'seed' is a random number which is used as the basis for determining the initial state of the diffusion. If you use the same prompt and default settings the image will be the same
1
8
more replies
Replying to @TomLikesRobots
((((((((((((((((((((((((((Yes)))))))))))))))))))))))))) I've tried...
1
2
Lol. I think the effect maxes out at around 120 pairs of parentheses but that could have just been the image I was working on.
1
1
more replies
Replying to @TomLikesRobots
Do you find repeating words as weight? Vampire, Vampire?
1
2
Not that much difference if that's all you do. I think if you rephrase it so have vampire portrait, lady vampire it reinforces it but it also changes the whole picture so you can't really do it incrementally.
3
Replying to @TomLikesRobots
Apparently not using parenthesis turns the vampire into Keri Russell
1
Replying to @TomLikesRobots
Quite the discovery, very practical too 👍
2
Replying to @TomLikesRobots
lol I love this. As if each set of parentheses tells the AI "seriously.... it's just a sidenote!"
1
Load more