I removed the extra memory cleanup step, as the single one is perfectly sufficient. I then added a dedicated LoRA section for each prompt section, allowing for targeted LoRA application to achieve better results. All descriptions can be found within the workflow.