Testing the krea2turbo merge to go from 8 to 4 stages – it’s not perfect but it’s a good start. For the trolls, a few glitches remain, so give it a miss – there’s no point getting worked up in the comments. Anyway, I won’t be replying to any hateful comments, as trolls are generally the ones who can’t get it together – you only have to look at their profiles to see that.
Description
FAQ
Comments (6)
Interesting model. Did you apply the Lora straight onto the fp8 or to the bf16 and save as fp8 if you don’t mind me asking? My results weren’t as good with euler/simple. I needed to go to 6 steps to get acceptable results.
converted directly on the FP8
This is interesting! In my (limited) tests - I was finding many images being under baked at 3 steps an overbaked at 4. Though, I have a lot more testing to say for sure. If dialed in right, it can be nice improvement for GPU poor folks.
In my tests on a 3070ti 8gb, it ran just as fast at 4 steps as an 8 or 12 step model with a low strength addition of the Turbo LoRA (and didn't perform as well either) - granted, it's faster than the 8/12 step :D - Not tying to say this model is bad, just noting findings to help make it better.
I would love to generate more things MUCH quicker.
Note: I was comparing to int8 models - so, this one may be faster if they were all same quant :)
Thanks for the feedback. I’m not convinced either – I was just giving it a go.
@Vince_AI Totally! Great to experiment, I am always happy to support faster generation and lower resource use :D