This LoRa was obtained by subtracting the MOSS-Image base model from the supervised fine-tuned model. You can apply it to the base model to bring its output closer to the fine-tuned version, although with better prompt following and less slop-like appearance in my opinion.
Use weights between 0 and 1.
I personally use the 8 step LoRa https://huggingface.co/lodestones/Kromoss/blob/main/8-steps-dmd2-lora.safetensors
and got the best results by first denoising a few steps with the base model only, followed by a few more steps with the SFT LoRa applied to the base model. See the example image for the workflow.