Using a lower a lower denoising strength with multiple loops can sometimes result in something that more resembles the initial image, but adheres to the prompt than if you simply did one pass with a a higher denoising strength. I find it to be a little tricky to use so I usually just avoid it and use controlnets etc. instead.
oooh now I see the conundrum. You are basically trying to find a way to sort of bake openpose into the image generation process… yeah I’m not sure how to do that other than what was suggested below or coding an extension for it.
Isn’t there a starting step/ending step parameter you can just use on controlnet with Autmatic1111? It uses percentage of the generation so just set the start to .5.
Pretty interesting, if I’m understanding it correctly though it is only out for Unbuntu and fairly hefty GPU (15GB VRAM).