This is a 1 image LoRA Workflow to generate a LoRA. Please place your settings where I have them set, as to make sure not to damage any of your hardware. This is safe, and I have made sure that the settings generate without too much stress on the system, namely the CPU if that is the only option you have to use Comfy. The Diffuser Model, Clip Models and Vae Model are the components of the SDXL Base Model; all of them together comprise the entire model. You can use just the model part of the Base Model to tie into the Lora without having the inference of the entire model; just one way that this can be used; I will let you figure out the rest.

*Revised*

If the process does not move very fast and you are using a CPU, there is one thing you can do to get a completed job no matter what. If you set the Learning Rate higher, the resources required to generate a LoRA will be lower so the process will speed up. So try, .0006000 or .0007000 or .0008000, slightly increasing it at that interval until you can complete the LoRA without your equipment getting too hot; since you reduce the Learning Rate, there is less demand put on your system, but, the higher the Learning Rate the less trained the finished LoRA will be in correlation to the data you used to create it.