Thats weird because flags are so simple How many times are you going to post this webm
Ethan Parker
Can you do the same on Euler_a?
Jacob Nguyen
my first and second attempt grids. seems simple enough if you just repeat the mantra "busty slut yoda" enough times
thanks, i haven't used the other UIs in a while so I wasn't sure about their implementations but it's nice to know i'm not missing out on that by sticking with voldy. I tend to just rearrange prompts rather than mess around with the weightings often anyway. might be kind of nice to have an option that randomly sets the strength of every token to a value between 0.9 and 1.1 on every generation of a batch, so once you have a decent prompt you can generate a very large batch and have a bit more variance within the output set
Codeformer has what they call a "quality" setting you pass it on the command line, which in reality appears to be a "level of bullshit" setting
so I give it a grainy blurry old photo of me the first larger pic on the far left is with quality "off" which means a setting of 1.0 on the command line (the numbers are reversed for some reason) It's good, looks exactly like me, just a clean and unblurred fixup of the original shitty photo
Then the others are with "quality" at 0.7, 0.5, 0.3 and 0.0 respectively (0.0 is full strength)
You can see at each step it's trying to make me more handsome, putting on makeup and making my eyes prettier etc. lmao
And amusingly the default "quality" setting if you don't specify isn't 1.0 (off), it's 0.5. So I guess they figured people would prefer the AI to bullshit them a little
I think you've misunderstood this webm, as per points out. E.g. 4 seconds into the 15 second clip (so ~40 steps) it's still mostly a blurry mess, and the final details aren't really done until 100-120 steps. Compare with the image I just made testing; at 40 steps the image is just barely perceptibly a little less crisp and finished than 150, nothing like the fuzzy mess at 40 in the webm. Feel free to run your own test and check. The benefit of running at something like 20, 30, or 40 steps (the lower end of that in euler_a mainly) is that you can very quickly generate a grid of 9 or 16, get a good idea of the average effect of your most recent prompt change, tweak something else, get another grid of 9, etc. You can do this a dozen times pretty quickly, then generate higher step outputs for the seeds you want to use as final outputs.
I ran out of yodatits, cooking some up right now >4000056822 woman with perfect closeup breasts and yoda head, green skin >4102421078 female yoda with huge tits, naked, closeup_breasts these are very technical terms, I don't expect you to understand >probably DPM2 20-50 steps
what's denoising strength in img2img do? seems to have the largest effect
Adrian Brown
"Yoda with big tits, busty slut, Yoda with busty green cleavage, plunging neckline, slutty yoda, green breasts, Star Wars concept art by Frederic Edwin Church, Michael Garmash",3447342692,512,512,Euler a,10.5,30,1662536149826-1.png
who knows maybe it really was in the concept art as an alt design at some point