Grey cat generated with the prompt "cute grey cat <lora:more_details:1>",
with "DPM++ SDE Karras", and "1" sampling step !
Stablity-Ai, the creator of StableDiffusion and many other things, released recently SDXL-Turbo, a generative model to generate pictures from a text prompt. The difference of SDXL-Turbo being that it is supposed to be much faster than SDXL-base.
So, i found a tutorial (on YouTube) on how to use this model with the interface ComfyUi, but i didn't find any convincing instructions to use it with Automatic1111.
So by looking here and there, I've finally understood why SDXL-Turbo is faster to generate pictures out of a text prompt.
Have a look at this page and read the PDF paper that is linked.
Of course, at the beginning, you have to download the model and put it in the usual directory.
First of all, it is created to make picture using only up to "4 steps".
By steps, they mean "sampling steps":
So this means that this model is faster because he has to do these steps fewer times, while not making the other steps slower.
For CFG Scale, you get the best results if you keep it at 1 (if more than 2, the results are bad).
For negative prompts, they don't seem very useful. However you can still use some without any loss of speed.
Also, when you use some, i am not sure it is even used at all.
Dimension: 512x512
For the sampling method, the "Euler" one, seems to work, but while some others may clearly bug, but "Euler" is not the only option.
You can actually use some Lora (once it is loaded in the memory, the pictures are then fast generated).
However, I think that the Lora has to be compatible with SDXL Turbo and while not uncommon, it is still a bit rare.
I've to say, I find the results pretty convincing and you can probably do a lot with this model.
Edit: I experienced some kind of bug while i started Stable Diffusion with an SDXL turbo model as last loaded / default to load model. It gave very blurry pictures while my settings were perfectly acceptable. The workaround seems to be to load another model, then you load the turbo model again. I wasn't able to reproduce the bug, so i don't know the exact conditions to make it happen.