This workflow uses Four KSampler approach. it does 4 steps high, then 2 steps high with Lightning lora, than 4 steps low, then 4 steps low with lighning. total 12 steps.

You may change the prompts. you need Q5_K_M text encoder and models for this to run on 12GB vram - I use Nvidia 4070 Super takes me exactly 600 seconds to do 12 steps like in the workflow settings