There have been a whole bunch of papers recently with optimizations that have led to new nodes in ComfyUI. I tested a few of them and I'm not that impressed, unfortunately. I used the Limitless Vision v4 checkpoint with the prompt: "a photograph of a woman cleaning a dirty apartment, wearing a french maid outfit, dusting"
A few VERY preliminary observations: PerpNeg and Perturbed Attention Guidance had pretty subtle effects, but increased coherence somewhat. ELLA is very aesthetic but actually decreased prompt adherence. I played around with different prompts a bit and could never actually get the apartment to be dirty - could be the checkpoint's fault. AlignYourSteps was the most promising. It seems to increase prompt adherence but decrease the coherence a bit.