Full size images found in this post.

This is a small article sharing results of a few tests I run to check what would be the best sampler settings for Stable Cascade.

Samplers & Schedulers

In all tests, Stage C and Stage B are run with the same sampler + scheduler pair.

Best:

• euler ancestral + normal

• uni pc + normal

Runner up:

• dpmpp 2m gpu + karras

• ddim + normal

• dpmpp sde gup + karras

Bad:

• dpm 2 ancestral + normal

• dpmpp 2m sde gpu + karras

• dpmpp 3m sde gpu + karras

Steps & CFG

Tests run with UNI_PC+normal.

Stage C

Stage B is run at 10 steps and CFG 1.1.

I personally think that 20 steps and CFG 6.0 gave the best results but there are many variations that were good.

Stage B

Stage C is run at 20 steps and CFG 6.0

Stage B was mostly influenced by its CFG.

I ran a second test with 8 steps (not 0.8) only changing the CFG.
I found that 8 steps and CFG 1.5-1.6 were giving the best results.

C & B CFGs

This test is run with Euler Ancestral + normal.

I wanted to see the interactions between different CFGs at stages C & B.
Results were not that conclusive.

Takeaway

The recommended settings for Stable Cascade are of 20 steps at CFG 4 for stage C and 10 steps at CFG 1.1 for stage B.

From these tests I found that I prefer to run 20 steps at CFG 6 for stage C and 8 steps at CFG 1.6 for stage B.

Stage C (compression) does have the strongest effect on the final outcome. That's the stage that is most open to experimentation.