LCM-LoRA merged with blue_pencil-XL
generate several times faster than usual, at a slight sacrifice in quality.
In my environment, a 1024x1024 image can be generated in 1 second. (5 steps, TensorRT, RTX 4090)
You must always follow these settings
Sampling method: DPM++ SDE Karras
or DPM++ 2S a Karras
or Eular a
Steps: 4 -
CFG Scale: 1 - 2.5
CADS, FreeU or other extensions may not give good result. (Increase the number of steps and it will stabilise.)
I don't know anything about LCM, Stable Diffusion or A1111, so I can't answer any questions.