Official Normal map compositing using the surface gradient framework in Shader Graph

This is great.
I read an article on normal blending, this article was going into performance consideration, I don't see your algo, how does it compare?
https://blog.selfshadow.com/publications/blending-in-detail/

 

So I have implemented this into a number of large shader frameworks now, and maybe I'm missing something or doing something wrong, but I find the visual difference very hard to spot in most cases.

For instance, toggling between regular tangent space normal maps and surface gradients on a basic terrain shader with height blending is visually identical. I suspect this is because the blend areas are small between textures due to the height blending. However, even when doing a linear blend it's not noticable.

Adding triplanar to the mix, the difference is again visually identical, which is surprising to me.

The one place I do see a difference is when layering normal maps - like detail texturing, or using macro normal maps to break up tiling. Here, the normal direction is a bit more muted and doesn't blow out as easy, but again I find the difference relatively minor.

In all of these cases, I'm converting from a tangent space normal map into a surface gradient using the interpolated TBN matrix, rather than say storing derivative maps directly, and I'm curious if this is causing my results to be generally the same?

 

I've been playing around with the surface gradient framework and the photoshop action from the Doing More With Less Unity blog post. If I convert the cliff normal map texture using the photoshop action the results look the same as using a tangent space normal map in a standard shader. However, when doing the same with spherical normals instead, the results are wrong.

If I use the original tangent space normal map of the sphere and convert it to derivatives in the shader, the result looks as expected. Is this an issue with the photoshop action?

 

Never mind, doing the normal map conversion in a shader and saving the result gives the same issues. So it seems to be due to the lack of texture precision. After observing the example image in the 'Doing More With Less' blog post, similar issues are noticeable there as well.