Thanks, they are 2D but they are created as 3d model so quality of their normal map is high. Sadly we wouldn´t be able to do per pixel destruction and splitting on 3d model, but we still wanted high visually quality, thus we chose this approach.
(note, they are some pixels around edges with wrong values, that will be fixed in clean up phase)
As I mentioned in different topic I am using modified Schlick's approximation, thus roughness maps instead of specularity + glossiness (which was important for both quality and optimization as far as Vram goes). Also we have pseudo screen space reflections (as far as they make sense in 2d)
this is kinda old gif but you get the idea.