SDXL. Speedup on default workflow, no speedup if you use lora.

[Ph0rk0z]

Added FBC to my normal SDXL workflow. Didn't notice any speedup. Used the demo workflow and there is indeed a 2s difference.

Only thing I'm doing is loading LoRA over the SDXL model. I tried adding the node both pre and post but the times don't change.

I'm on linux with latest stable torch and using xformers attention. Set the value to .2 just like the demo. What am I missing?

[blepping]

Are you using an ancestral or SDE sampler?

[Ph0rk0z]

No, but I found the scheduler is the culprit. I went through his workflow piece by piece and found the main difference is it uses exponential. I use sgm_uniform, generally that's my favorite. Simple is out too. Will have to go through all of them.

[blepping]

#87 might be of interest to you - has some research on how different settings affect quality/speed. (I don't believe there was any testing with different schedules yet though.)

[Ph0rk0z]

That helps for flux but XL will be different. Issue is that all the "good" schedulers don't work with this. Either it messes up faces or it provides no speedup. On some turning down the threshold helped. GITS was the most interesting because it really turbo-ed the generations.

I think I know why on flux I had no help from TeaCache either. Nobody has brought up how schedulers affect this so far. IMO, should be a warning to not confuse people.