This workflow is designed to be used by Merged SDXL checkpoints as well as SDXL LoRA. It should work well with previous versions as well, I simply don't use them however. Both the original render as well as the refined image are upscaled and saved in dated directories under the Upscaled tree. The reason I save the original in upscaled format is that on occasion the facial refiner will sharpen and/or distort a background face that is blurred out. By saving the original, I can go into Photoshop and set the original as the bottom layer and the refined as the top. Then I can erase that section of the refined image before merging and saving the final product. After I have the final image, I upscale it further using Gigabit AI before sending it off for poster printing In the KSampler, I perform an increment of the seed. This lets me rewind in case there's something unusual about the face refinement, and thus produce a new face for the same image. Also, this workflow only works with checkpoints that have merged the VAE into themselves as there is no seperate VAE selector as well as no SDXL refinement pass.