New paper on inversion. (Finding the noise from the image)

[sharp-trickster]

https://arxiv.org/abs/2211.12446

The idea of interrogating both the prompt and initial noise that generated some image, then editing details seems like a more reliable way to your prompt changes have the intended consequence on the new image, as a pipeline. There must be more interesting and useful stuff that you can do with it too. The paper examples look promissing.
There´s no code on this one yet, too, sadly, but who knows, at the pace we get new amazing stuff published and implemented.