I've seen these same FSR issues in every game. XeSS does seem to vary in quality more often.
Normal behavior when cpu limited all 3 have a fixed cost so using them when cpu limited usually has a performance penalty.
Just a tradeoff it seems. FSR doesn't need any special instructions since it just runs on shaders.
XeSS needs some form of support, and just like Nvidia it needs XMX to perform the best.
DLSS is proprietary and works best on Nvidia GPUs with the best tensor cores.
By the way for XeSS, we're getting another hardware with fundamentally different performance levels. On Meteorlake's XeLPG, it doesn't have XMX so it's not up to their dGPUs, but keeps the concurrent FP+Int/EM execution, so it's a faster version of DP4A then on the XeLP of Tiger/Alder/Raptor.