This uses the AlignYourSteps scheduler and Perturbed Attendance Guidance nodes as part of the image generation with Face ID and a ControlNet Depth map using MiDaS for the preprocessor, a face detailer, 1st stage Upscale using SUPIR and 2nd stage upscale using model.

It's spilt into groups and there's a Fast Group Bypasser. The only group you can't turn off is the Txt2Img group. The rest can be tailored to what you want to use.

Starting from top right, the first group (yellow), has a standard KSampler, so you can run a comparison of a gen without AYS to the main one. It has an image comparer hooked up to the non upscaled AYS gen for easy comparison and a save node.

Second group (black), is for FaceID.

Third group (pink), is for ControlNet.

Fourth group (purple), is the Txt2Img group. It has a single seed control, so if you want to run tests with a fixed seed you only need to set it on the node in this group. It also has Fooocus Styles selector node and a couple of CR LoRA stacks.

Fifth group (teal), is the face detailer group which has a save node for the non upscaled version. There's a separate positive prompt box for faceial features, so say you're trying for blue eyes but they keep coming out brown, you can prompt blue eyes in there to correct it.

6th and final group (blue), is the 2 stage upscaler.

I get more CUDA errors than successes on it, although I think it's either an issue with my OS/System or my 2080Ti with 11GB VRAM isn't enough. I have had it complete enough times to be confident that the workflow is fine. If you're on similar or less VRAM to me you may want to consider bypassing the upscaler group and doing that task in a separate upscaling workflow.