Realism Enhancer is a full LoRA, not a slider. It was trained exclusively on high quality, 2048px plus realistic images, with the goal of pushing the model toward real photography, natural skin, and detailed textures.
During testing, it was able to break MOSS's tendency to generate cartoon or anime results in several cases. It is not perfect, and results will vary depending on the image and setup, but the training seems to help it move toward realism. When the image is already realistic, it can improve skin and textures considerably.
I recommend starting at strength 1, but feel free to test different values to find what works best for your setup. I tested it with Euler/simple and got good results, though ER_SDE/Bong_Tangent worked even better in my tests.
All example images have the workflow embedded, so you can load them in ComfyUI, inspect the settings, and reproduce the results yourself.