@TzigoAI Bit late, but for anyone confused about this: Because it is a Hyper CFG version, meaning that it works with CFG greater than 1, which activates the negative prompt. So both positive and negative conditionings get diffused per step. It's how CFG works. Hence twice the duration. It's the same for both CPU and GPU.

The non-CFG versions (1, 2, 4, and 8) are supposed to be used at CFG 1 (thus disabled) and only diffuse the positive conditioning. However, if you'd set CFG higher for them they'd also take twice as long.

@Sajrep i'm a little confused. So for the 12 step CFG version will required setting CFG between 5.8 with 12 steps??

@PinkywDreams Yes, though the recommended CFG range is 5-7. You could also try more steps if 12 is not refined enough. It varies per model that it gets applied to.

@Sajrep If I want to apply it on Boltning Hyper SDXL checkpoint model. which one would you recommend? Coz it requires CFG scale set to 1...

@PinkywDreams Boltning HYPER already has the Hyper distillation built-in. Probably Hyper8 (non CFG) hence the CFG 1 requirement. Adding Hyper12 CFG on top of that won't work. Also won't work with the Lightning versions as it's a similar modification.

@Sajrep i think i got the idea!! thanks man

@Sajrep Sorry for replying to an old thread, but I started using these Hyper Loras, Non-CFG and CFG, and on the Non-CFG the time it takes with any CFG (1 - 4.5 what I have been using) is literally the same and the rendered result does change considerably from 1 to 4.5. I have not been able to generate good renders from the CFG Version Loras, I get results that look like a painting was washed down with water, smudged all over. I am also using DrawThings, so I don't know if that is why I am getting these results (speed and CFG versions not working). CFG Versions do take longer though. I also have been using NON CFG versions with CFG in 1 while changing noting but the Negative prompt and I do see changes in the result, so I don't know if it really ignores the negative prompt or, again, it has something to do with DrawThings.

@maloofmx I can't really vouch for Draw Things, but I did test it a little and non-CFG Hyper did work as expected on the version I tried. I did find it finicky though, like you can easily accidentally do img2img instead of txt2img because it tries to be "smart".

The idea is that CFG is turned off by setting it to 1. Negative/unconditional doesn't have an effect at that point and can be skipped. This is inherent to the diffusion app, not something the LoRA does.

Just to be sure: It's important to use the correct models, so SDXL Hyper with SDXL checkpoint and such, otherwise the LoRA would be ignored. It also depends on the model how well it works.

Generally when total steps time required is disproportionate, it is because you have exceeded your gpu (or iGPU) memory cache, and ram begins to be used excessively. You could try 12 steps without the lora and compare its speed, it should be the same. If you can enable tiling in your UI you use, it should help.