Open unphased opened 1 year ago
So far what I'm able to do to get reasonable throughput is to set my sdxl model paths correctly, then prepare a prompt, and then save the workflow, and then i can repeatedly issue the generation by pasting each time before generating. This is obviously far from ideal.
Edit: it looks like using Load from history also works. In fact somehow once restoring from history it seems to be working properly. Not sure what's going on.
Thanks your testing out my workflow! so the refined image never actually touches the base image. how its set up is as follows:
Prelim > Base Prelim > Refiner
The preliminary image is 80% complete, and then I finish it off twice, once with the refiner and once with the base. the last 2 steps can complete in any order.
Looks like with the first generation it works right, but in subsequent ones it goes straight from preliminary to the refiner and then spits out a base output. The quality suffers because only the first ever generation has the proper 3 step process. These messed up subsequent runs also complete a lot quicker than the first run.