My recommended settings
- LoRA Strength: 0.9 (or possibly even lower)
- Steps: 16
- Sampler: DEIS
- Scheduler: KL Optimal
- Shift: None (I removed the node, since it made no difference after I swapped to KL Optimal scheduler.)
Reason for making
The 4-step LoRA does fairly well at 4 steps, but it cannot go higher than 4 steps without overcooking the image, and even at 4 steps the image feels a little cooked.
4-step lora | 4 steps vs. 8 steps vs. 16 steps | [1536x1536, no shift, lora strength 1, deis, kl_optimal, seed 187]
The 8-step LoRA on the other hand is very undercooked at 4 steps, still a little undercooked at 8 steps, but handles higher step counts like 16 really well, but feels a little undercooked overall.
8-step lora | 4 steps vs. 8 steps vs. 16 steps | [1536x1536, no shift, lora strength 1, deis, kl_optimal, seed 187]
Merging these two together results in being able to do 16 without overcooking or undercooking. It feels just about right, especially if you load at 90% strength.
merged lora | 4 steps vs. 8 steps vs. 16 steps | [1536x1536, no shift, lora strength 1, deis, kl_optimal, seed 187]
