Curiosity with visual observations?

Hi, i read somewhere that is not right to use curiosity rewards with visual observations because the predicted image is noisy. The explanation was that curiosity predicts pixel by pixel. Is that right? Is not good to use curiosity with visual observations?

 

AFAIK, a possible issue here is that if visual observations are particularly noisy at some point in the environment (like an in-game TV screen), the curiosity reward can spike there, causing the agent to get stuck. I don't think you need to worry about overall pixel noisiness though. I'd assume curiosity rewards correlate to features extracted from visual observations, rather than to individual pixels.

 

Ok, thanks.
Another question: im not using curiosity in my training, is my agents performance guaranteed to improve if i use curiosity? i.e. is always a good thing to use curiosity? If is not, when is it better to use it?