Earlier quoted context omitted.
Maybe they meant 50%-200% slower, in which case the x-factor range would really be 1.5x to 3x?
200% slower is 3x as long, yes.
Running Stable Diffusion XL 1.0 in 298MB of RAM
161–166 of 166 posts
Re: Running Stable Diffusion XL 1.0 in 298MB of RAM
#162Earlier quoted context omitted.
What? No, that's confusing enough it's almost hostile. The fact thst your math is wrong is proof enough. 50% to 200% more time is 1.5x to 3x slower. I don't like how it was worded by the author. But all you've done is essentially invert the wording while making the math MORE difficult in the process.
Umm what? I hear “50% more time” all the time. 50% to 200% is 0.5x slower to 2x slower. People seem to be confusing “% slower/more time” vs “% of current time”
I think something like “0.5x slower” is just conventionally not used, even if it’s understandable. It’s a difference in common usage between percentages and x-factors. One reason might be that x-factor implies multiplication, not addition; that’s what the “x” stands for. X-factors are typically used to say something like the ‘the new run time was 1.7x the old one’. Whether it was slower or faster is implied by whether the x-factor is below or above 1.0. Because x-factors are commonly used as multipliers, and not commonly used to say “1.5x more than” (which actually means 2.5x), it’s pretty easy for people to misunderstand when someone says “0.5x slower” because it looks like an x-factor.
Now, the same argument could apply to percentages. A percentage is also a factor. But in actual usage, “50% more” is common and “0.5x more” is not; and “100x faster” is common (usually to mean 100x not 101x) while “10000% faster” is not common at all. So language is inconsistent. ;)
All that said, using an x-factor as a pure factor, and not a multiply-add, is less confusing and more clear. Saying “The new runtime is 1.5 times the old one” leaves no room for error, where “The new runtime is 90% slower than the old one” is actually pretty easy to miscalculate, easy to mistake, and easy to misinterpret. The percentage-add is also asymmetric: 90% slower means 0.1x, while 90% faster means 1.9x. Stating a metric as percentage-add makes sense for small percentages, and makes more sense for add than subtract once the numbers are double-digit and larger.
Re: Running Stable Diffusion XL 1.0 in 298MB of RAM
#163Earlier quoted context omitted.
This is the best I've tried so far, but no mac support I don't think. Its a feature packed fork of Fooocus, which was developed by the orginal ControlNet dev. The quality you can get from small prompts is mind boggling: https://github.com/MoonRide303/Fooocus-MRE For base SD 1.5, I use Volta, because its fast: https://github.com/VoltaML/voltaML-fast-stable-diffusion/com... Really good SD 1.5 image quality comes from g…
Fooocus does quite a bit of prompt massaging for you - there are models that take a few words and turn them into “prompt engineer” level prompts. Makes a huge difference.
Still, even with it turned off, the quality is quite remarkable.
Re: Running Stable Diffusion XL 1.0 in 298MB of RAM
#164Earlier quoted context omitted.
200% slower is 3x as long, yes.
But 50 % slower is either 1.5 or twice as long, depending on whether the 50 % refer to an increase of the runtime or a decrease of throughput. 200 % slower is only unambiguous because the throughput can not decrease by more than 100 %.
I’ve been doing capacity planning lately and the whole deal with how we are bad at fractions came up again. If you have to spool up 10% more servers and then cut costs by 10%, you’re still slightly ahead. If you cut 10% and then later another 10% you have cut 19% of the original, not 20%.
Re: Running Stable Diffusion XL 1.0 in 298MB of RAM
#165Earlier quoted context omitted.
Fooocus does quite a bit of prompt massaging for you - there are models that take a few words and turn them into “prompt engineer” level prompts. Makes a huge difference.
Yeah, and InvokeAI has a similar "IP-adapter" model. Still, even with it turned off, the quality is quite remarkable.
Ip adapter uses an image to guide denoising.
Fooocus and MJ take a prompt and expand it in a variety of ways (eg a language model or more simplistic text manipulation). The actual prompt that creates the conditioning is not what you typed in. That’s what I mean by prompt massaging
Re: Running Stable Diffusion XL 1.0 in 298MB of RAM
#166Earlier quoted context omitted.
https://en.wikipedia.org/wiki/Wikipedia:Two_times_does_not_m...
This is an issue that irritates the hell out of me when I see it and I'm so glad this person wrote it out so plainly.