Driver RIS and FidelityFX CAS use the same technologies. The reason it's also available for the developer to integrate, rather than being driver only, is that the driver implementation necessarily can only apply it on the final presented image the game sends for display. That often looks fine, but it might not be optimal depending on what's being drawn and how it's being drawn.
The best case is the developer integrates it into the right place in their renderer, usually somewhere in the post chain that correctly considers colour space, text and UI rendering, etc. If that happens and a game integrates it, the driver uses app registration and other mechanisms to disable RIS for that title, so it doesn't happen twice.
CAS integrations were some of the last things I worked on at AMD, and the feedback from developers was very positive on the whole. Where it didn't work in eval were sensible cases that tended to be technically related, rather than quality related. Things like the game already had a well optimised content-aware sharpener tightly integrated into an existing pass.
One of the key selling points, and a real feature, is the fact that it's completely open and MIT licensed, so developers can do what they want with it. Pull it to pieces, optimise it, integrate it however you see fit, use it as a building block for something else, whatever they like. No black box, no silly licensing, no silly deals. Just "here's some cool tech we invented that you might like and it's free".
