Tangent space computation is under-specified

[Themaister]

Reference: KhronosGroup/glTF-Sample-Models#174

When you have engine-computed normals, the spec says that the application should use the "default MikkTSpace algorithms", but I have run into issues where MikkTSpace and the reference code for generating tangents do not match in the W component.

The issue which needs to be clarified is:
What is the inferred bi-tangent sign when loading a model which does not have tangents? The reference PBR shaders seem to suggest that W is always 1 (i.e. B = cross(N, T)), but that disagrees with MikkTSpace output in many cases.

In general, I would like to see tangent-generating better specified when TANGENT accessors do not exist.

[Ybalrid]

Tehre's also the issue that even if you have normal, you may not have tangent and/or need to recompute the tangent vectors. This redirect you to Appendix A, that is still a TODO.

Should something about recalculating TANGENT vertex attribute, because there's an infinite number of tangents to a surface. All I know is that common practice is to align TANGENT and BITANGENT (BINORMAL) with the UV plane.

[emackey]

The glTF spec does specify the format of the normal maps though:

The normal vectors use OpenGL conventions where +X is right and +Y is up. +Z points toward the viewer.

I tried to spell out the implications of that in this comment: mrdoob/three.js#11438 (comment)

[Ybalrid]

Thanks @emackey, I will read through that carefully when I'll have some time! 😄

[lexaknyazev]

When tangents are missing, the engines should generate them using MikkTSpace algorithms based on the positions, normals, and texture coordinates present. Bitangent vectors are always computed as a cross product of normals and tangents multiplied by the tangent's W component.

Any issues with particular implementations should be reported to their repos.