Here is another example. Comparison of sampling quality of different approaches.
Models: Chroma unlocked, v. 48, float 8, e4m3fn, scaled, calibrated. Flux furry enhancer, LoRA
Euler sampler, beta sheduler:
- First pass, 8 steps, 3.5 guidiance.
- Second pass, 16 steps, 7.0 guidiance.
- One pass, 32 steps, 3.5 guidiance.
- One pass, 32 steps, 7.0 guidiance.
As you can see, two-pass approach, 8+16=24 steps, bring very good result in comparison with one-pass 32 steps result. Not so good, but adorable. But if we look closer, we can see drastic difference. There is a fur, or tiny hair, on the breast of the "two-pass", 24 steps, image, while one-pass demonstrate good details, but kind a flat and unrealistic. Risng steps further do not change that issue. Euler seed is almost still behind 26-32 steps.
Last image demonstrate completely flat and cartoonish look, but very sharp details, cause of high prompt tension/guidiance. So, if you want to add details to image - use it on top of base image. For regular use, and more realistic look, try lower value. And the last. Amount of steps also generate variance of image. It can be seen especially with dpmpp_sde_gpu and lcm sampler. Keep it in mid - use very detailed and strict prompts to keep overall image look under control. Hope it was useful. CYA.