Normals should be multiplied by the transposed inverse of the transform matrix.

[HackerFoo]

I noticed that normals are not transformed correctly when scaled non-uniformly:

bevy/crates/bevy_pbr/src/render_graph/pbr_pipeline/pbr.vert

Line 30 in f4776f2

v_WorldNormal = mat3(Model) * Vertex_Normal;

https://www.scratchapixel.com/lessons/mathematics-physics-for-computer-graphics/geometry/transforming-normals

Example (a sphere stretched along one axis) with incorrect normals:

Corrected normals:

I used this as a quick fix, but computing the inverse in the shader isn't a great idea:
HackerFoo@a711633

The proper fix would pass the precomputed transposed inverse into the shader.

[jakobhellermann]

This is already fixed in the pipelined-rendering branch:

bevy/pipelined/bevy_pbr2/src/render/pbr.wgsl

Lines 46 to 50 in 43e8a15

	out.world_normal = mat3x3<f32>(
	mesh.inverse_transpose_model.x.xyz,
	mesh.inverse_transpose_model.y.xyz,
	mesh.inverse_transpose_model.z.xyz
	) * vertex.normal;

bevy/pipelined/bevy_pbr2/src/render/mod.rs

Line 85 in 43e8a15

inverse_transpose_model: transform.inverse().transpose(),

[CptPotato]

There's also #2049 which I kept as a draft since a few questions regarding transform propagation and shearing haven't been resolved, yet. It's probably not relevant anymore, though, as it's for the old renderer.

[alice-i-cecile]

Adding to 0.6 milestone; this should be closed out once pipelined-rendering is merged in.

[Weibye]

@alice-i-cecile Since pipelined is merged, this can be closed right?