I understand this allows me to ask any rendered pixel what contributed to it's color in detail?
IIUC, yes, but even more so - it allows you to fit a rendered image to an existing one by doing a gradient descent along the pixel errors along the parameter space.