Open Butterfly-Dragon opened 10 months ago
Something to note: before writing here i noticed in the bugs that people seem to complain that SDXL Turbo does not load properly if one does not load SDXL before running it. I see it loads itself with the SD 2.X inference yaml.
That is unrelated to this but might be worth checking.
The speed of SD XL Turbo is (indeed) the same as my SD 2.X models. But the results are eons better. Also the results are eons better than standard SDXL too given a "same steps count". It is just incompatible with SD 1.5 loras but at this speed the price is one worth paying.
im not at all savvy at this stuff but is there any way to speed up the generations? im at like an hour per gen
@ryanoZphoto what were you doing besides generating, and what model type did you use?
So far what i noticed is that:
Resolution is the second most important thing, as the speed scales with the megapixels, hence why most people prefer to upscale rather than generating at high resolution.
a 512x512 image (0.25 megapixels) is around 16x faster to generate than a 1024x1024 picture (1 megapixel) so you go from 2 iterations/second to 1 iteration every 8 seconds.
Next thing i saw that accounts for that is what else you do with your machine when generating. You get the fastest results by closing everything and letting the machine churn. Yes, that includes the browser with WebUI, as it would otherwise generate previews which further slow generation. Basically, start the project and if you are not in the "fiddling fase" where you are chosing what to ask and how, just let it do its thing. This is impacted by model size. Smaller models (2 GB) are less impacted by what you do with your machine, but if you have to load a 9GB model then even breathing near your machine will slow things considerably.
Obviously this is comparing a machine to itself, you said absolutely nothing about your machine, what were you doing, the size of your generation so, therefore your comment is pretty much useless, so i tried to give broad strokes to address what might be the problems you have and how to work around them according to my empirical personal experience.
I had been on v1.5.1 and recently got back into it after a break. I updated to v1.7 and noticed the same extreme slowness.
When I tested, all things being equal and using the same exact model (making a copy in windows), 1.5.1 yields considerably faster results than 1.7. I tried several models, some based on SD 1.5 and some based on SDXL all came back much faster in A1111 v1.5.1. I even have my v1.5 copy on a standard HDD. I tried installed 1.7 on my nvme drive and it made no noticeable difference.
For now I'm sticking with A1111 v1.5.1 simply because I know it works, but I'd really love to use the new features in v1.7. If anyone has any insight into what's going on with the speed I'd love to hear it as well.
Checklist
What happened?
Stabe diffusion XL averages between 45 to 60 seconds per iteration on a 512x768 image Stable diffusion XL turbo averages between 3 and 4 seconds per iteration on a 512x768 image Most SD 1.5 models average around 1 to 2 iterations per second on a 512x768 image
My system specifications:
Steps to reproduce the problem
What should have happened?
Most people i see seem to have "iterations per second" not "minutes per iteration" (if i start watching a video on youtube or tiktok or use the graphics card for anything the "seconds per iteration" can easily get in the hundreds) even on a base SDXL model and the comments i read state something about "waiting a couple of seconds" as the difference between SDXL Turbo and SDXL base model.
What browsers do you use to access the UI ?
Google Chrome
Sysinfo
sysinfo-2024-01-29-14-12.json
Console logs
Additional information
nothing else to add. But for the sake of completeness let me add what i was doing above. The following are (in order):
imposing tall adult perfect, tasteful masterpiece
, 150 steps (reduced because img2img) CFG 10 seed 1096601817