well it's 2020 now and I still agree with this

It's quite tricky to use parallax occlusion mapping. First - it's very slow, even on modern hardware. Many developers like simple parallax with one or two texture fetch, because such approach is way much faster, but it's not so cool. You can see Quake 2 Berserker mod and you will see that this mod uses POM for many wall surfaces and can slowdown even GTX 680. :) Second - it not appliable for all models, because of artifact prone nature of algorithm. It actually very appliable on square surfaces like walls, because edge of surface can be edge of texture, they both square like form and hardware clamp to texture edge can be provided, so algorithm can't sample wrong pixels out of the edge pixels. In complex models like human, edges of surface are not related to texture edges so algorithm can sample wrong out of edge pixels, especially when model will be animated. Also to give good performance it make sense to keep height information out of the normal map, because parallax didn't required big texture to nice looking results and with big texture it way much slower. It's not actually a problem, but still not very comfortable approach. So that's reasons why developers didn't like parallax occlusion mapping very much. But it still very cool looking for a walls, floors, doors and lot square like objects. :)

@@MegaSmouke I think that Cryostasis use POM on almost everything. Even on complex models. Maybe thats reason why is it so demanding

Yes, in 2005!!!