Features
added SDXL and SDXL Turbo support (#406, #431)
new pipelines: SDXL txt2img, SDXL img2img, SDXL panorama
added DPM SDE scheduler
added LCM support for SD v1, v2, and SDXL (#421)
added ...
Are there any benchmarks available? I can’t find any in their Githup repo. I’d like to run SDXL Turbo txt2img on my old computer. Linux, 4 core Intel Xeon, 32GB of RAM and no graphics card, so CPU only.
How old of a Xeon? Because it won’t be a fast result, but maybe you are fine with it. Back when I tried this, SD 1.5 could do 20 steps at 512x512 on my Ryzen 5 3600 in roughly 7 minutes…
It’s a Skylake one, so from 2016 or something like that. I’ve tried stable-diffusion.cpp this takes like 15min for 20 steps on my machine, I’ve never measured it exactly. But it doesn’t have that much features.
What I’d like to do is use one of the 'Turbo’s. Just doing 1-4 steps sounds like it could make it close to usable. Guess I need to test it myself. Maybe I’ll have some spare time later today. I just wanted to avoid fighting with onnx. Last time I tried to use it, it was kind of a hassle to get it running.
Edit: Oh wow, I missed the BENCHMARK .md file in the repo 🙃
Are there any benchmarks available? I can’t find any in their Githup repo. I’d like to run SDXL Turbo txt2img on my old computer. Linux, 4 core Intel Xeon, 32GB of RAM and no graphics card, so CPU only.
How old of a Xeon? Because it won’t be a fast result, but maybe you are fine with it. Back when I tried this, SD 1.5 could do 20 steps at 512x512 on my Ryzen 5 3600 in roughly 7 minutes…
It’s a Skylake one, so from 2016 or something like that. I’ve tried stable-diffusion.cpp this takes like 15min for 20 steps on my machine, I’ve never measured it exactly. But it doesn’t have that much features.
What I’d like to do is use one of the 'Turbo’s. Just doing 1-4 steps sounds like it could make it close to usable. Guess I need to test it myself. Maybe I’ll have some spare time later today. I just wanted to avoid fighting with onnx. Last time I tried to use it, it was kind of a hassle to get it running.
Edit: Oh wow, I missed the BENCHMARK .md file in the repo 🙃
Got it working. My machine needs 35s per step. So 2-3 mins for an SDXL-Turbo image. And 20s/step for an SD1.5, so 7mins for the whole image
Can I offer what I believe is a better option? 1.5 LCM models. 5 steps for a good image, and they’re 5 steps at 1.5 speeds :)
I like this but obviously you can find other LCM models.
Awesome, thank you very much.