A reference for different samplers using Stable Diffusion

InvokeAI released a 2.0 version yesterday. After fiddling around with it, and
finally getting everything working on my M1 MacBook I wanted to do a comparison of each of the samplers for future reference.

I used the same prompt, steps, cfg_scale, and seed for them.

Prompt:

polaroid photo of high mountains, photorealism

Other Settings:

Side note: I set the steps low as I was limited on time, I will touch on that in the Final Thoughts section

steps: 20
cfg_scale: 7.5
seed: 3971012830

Image Output

image: 000004
Sampler:  ddim
1 image(s) generated in 53.50s

image: 000005
Sampler:  plms
1 image(s) generated in 63.51s

image: 000006
Sampler:  k_lms
1 image(s) generated in 64.39s

image: 000007
Sampler:  k_dpm_2
1 image(s) generated in 98.26s

image: 000008
Sampler:  k_dpm_2_a
1 image(s) generated in  99.59s

image:   000009
Sampler:  k_euler
1 image(s) generated in 55.66s

image:   000010
Sampler:  k_euler_a
1 image(s) generated in 56.86s  

image:   000011
Sampler:  k_heun
1 image(s) generated in 100.47s

Final Thoughts

k_euler_a, k_dpm_2, and k_dpm_2_a samplers really could use some additional steps. See below for results with 50 steps.

ddim and k_euler were the fastest to generate without too many steps.
They may be great for getting your initial image.

I was initially pretty surprised by how each sampler provided a really close
image structure to each other using the same seed.

I am also continuing to be surprised by how powerful a little M1 MacBook Pro
can be in this arena without tripping on the fans at all.

k_euler_a, k_dpm_2, and k_dpm_2_a samplers bumped to 50 steps:

image: 000013
Sampler:  k_euler_a
1 image(s) generated in 136.07s

Seems better, but not worth the extra time?

image: 000014
Sampler:  k_dpm_2
1 image(s) generated in 236.99s

Still blurry ?? No improvement really, just waste time.

image:  000015
Sampler:  k_dpm_2_a
1 image(s) generated in 236.18s

Still blurry ?? No improvement really, just waste time.