Derivative maps vs. Tangent Space Normal maps

Question

I discovered that some engines use derivative maps instead of tangent space normal maps.

After some reading, it seems to be a really awesome way to replace tangent space normals but are there some disadvantage using them ? Why still continue using tangent space normals ?

Is it possible to compare both with advantage and disadvantage ?

Answer 1

After some researches and some answers from professionals here is my conclusion.

Pros

• Don’t require tangents or binormals. Less interpolators.
• Only need two channels. less texture memory.
• Don’t suffer from tangent seams.
• Can be blended using alpha blending, without renormalization.
• Less mesh memory: We don’t need to store a tangent vector.
• Fast implementation.

Cons

• More ALU
• Less flexible. A normal map can represent any derivative map, but the reverse is not true. As an example, sharp edges can be difficult to represent.

So, that's a lot of pros compared to cons. But the major problem is that it's not an industry standard.
There are almost no content authoring tools nor artist know-how.

Here's is a quote from Bart Wronski that illustrates well the current status of derivative maps :

Sadly in life / technology not always best solution wins / not even gets deserved attention... It's more about standards and inertia.

If you are interested in knowing more about derivative maps here are some interesting articles.

• Mikkelsen and 3D Graphics: Derivative Maps | Morten Mikkelsen Blog
• Derivative Maps | CodeItNow
• Derivative Maps vs Normal Maps | CodeItNow
• Triplanar Texturing with Derivative Maps | The Robot Seven
• Derivative Normal Maps - What are they? | polycount.com
• Derivative Maps | Knald Technologies Documentation

If I forgot something or if your don't agree feel free to tell in the comments, I would be glad to improve this answer.

Answer 2

I assume that you're using precomputed height map derivatives rather than calculating them on the fly (for details see this post on Mikkelsen's blog). If we need to supply pre-computed height derivatives, then we have to supply two channels, just like a normal map. One could argue that derivative mapping doesn't require the presence of a tangent vertex attribute like normal mapping does, but the extra differentiation operations on the height map somewhat nullify that performance gain. Derivative maps are a cool concept, but at the end of the day I don't think they are significantly better than normal mapping performance-wise (although conceptually I agree that derivative maps are easier to work with because we don't have to deal with tangent space).