Confusion about --lowvram and --medvram

[Examiner1]

There's --lowvram and --medvram, but is the default just high vram?

[ClashSAN]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Optimizations

Makes the Stable Diffusion model consume less VRAM by splitting it into three parts 

there is no --highvram, if the optimizations are not used, it should run with the memory requirements the compvis repo needed.
There is an opt-split-attention optimization that will be on by default, that saves memory seemingly without sacrificing performance, you could turn it off with a flag.

[Examiner1]

Thanks!

[illtellyoulater]

@Examiner1 but what are the disadvantages when using lowram or medram?
Lower model loading speed?
Lower image generation speed?
Something else?

[CTimmerman]

--medvram Lowers performance, but only by a bit - except if live previews are enabled.

--lowvram | An even more thorough optimization of the above, splitting unet into many modules, and only one module is kept in VRAM. Devastating for performance.

[SSalvadoRR-official]

How much vram is actually low and med?

[CTimmerman]

How much vram is actually low and med?

--medvram with a 2GB model only uses 3 to 6 of 64 GiB RAM while generating a 60-step Euler 512x512 x1.2 = 614x614 image on my 6 GiB GTX 1660 Ti in 1 min. 45.1 sec.
I haven't tried --lowvram as even 6.62 GiB XL models work but take hours. Maybe I should as swapping might cause the slowdown.