The unskinned and skinned vertex shaders transform normals by the model matrix's upper-left 3x3 directly. The correct transform for a normal is that matrix's inverse-transpose. Any node whose world transform has non-uniform (anisotropic) scale or shear shades incorrectly: the normal points the wrong way, corrupting both analytic lighting and IBL.
Where:
flutter_scene_unskinned.vert:26: v_normal = (mat3(frame_info.model_transform) * normal).xyz;flutter_scene_skinned.vert:69: v_normal = (mat3(frame_info.model_transform) * mat3(skin_matrix) * normal).xyz;- Consumed at
flutter_scene_standard.frag:234 (normalize(v_normal)), also fed to IBL via env_normal at :268.
Why: positions and tangents transform by M, but a normal must transform by (M^-1)^T to stay perpendicular to the surface. mat3(M) equals (M^-1)^T only for rotation, uniform scale (up to a positive scalar that normalize removes), and reflections. It diverges for non-uniform scale and shear, a direction error that survives normalization.
Not a regression, and unrelated to the winding fix in #137: rotation, uniform scale, translation, and axis mirrors (negative unit scale) all render correctly, since reflections are self-inverse-transpose.
Repro: a CuboidGeometry + PhysicallyBasedMaterial (low roughness) under a Node with scale(3, 1, 1). Its specular highlight is displaced versus the same shape modeled at 3x1x1 with an identity transform.
Proposed fix: add a normal matrix to the FrameInfo UBO, computed CPU-side as transpose(inverse(mat3(model_transform))), and use v_normal = normalize(normalMatrix * normal) in both vertex shaders. Skinned path: same treatment relative to model * skin (rigid joints are already fine). Mind std140 layout (pass a mat4) and the web shim repacking, then rebuild the shader bundle.
Priority: correctness, lower than the winding fix. Many models bake non-uniform scale into geometry or avoid it, so real-world impact is limited.
The unskinned and skinned vertex shaders transform normals by the model matrix's upper-left 3x3 directly. The correct transform for a normal is that matrix's inverse-transpose. Any node whose world transform has non-uniform (anisotropic) scale or shear shades incorrectly: the normal points the wrong way, corrupting both analytic lighting and IBL.
Where:
flutter_scene_unskinned.vert:26:v_normal = (mat3(frame_info.model_transform) * normal).xyz;flutter_scene_skinned.vert:69:v_normal = (mat3(frame_info.model_transform) * mat3(skin_matrix) * normal).xyz;flutter_scene_standard.frag:234(normalize(v_normal)), also fed to IBL viaenv_normalat :268.Why: positions and tangents transform by M, but a normal must transform by
(M^-1)^Tto stay perpendicular to the surface.mat3(M)equals(M^-1)^Tonly for rotation, uniform scale (up to a positive scalar thatnormalizeremoves), and reflections. It diverges for non-uniform scale and shear, a direction error that survives normalization.Not a regression, and unrelated to the winding fix in #137: rotation, uniform scale, translation, and axis mirrors (negative unit scale) all render correctly, since reflections are self-inverse-transpose.
Repro: a
CuboidGeometry+PhysicallyBasedMaterial(low roughness) under aNodewithscale(3, 1, 1). Its specular highlight is displaced versus the same shape modeled at 3x1x1 with an identity transform.Proposed fix: add a normal matrix to the
FrameInfoUBO, computed CPU-side astranspose(inverse(mat3(model_transform))), and usev_normal = normalize(normalMatrix * normal)in both vertex shaders. Skinned path: same treatment relative tomodel * skin(rigid joints are already fine). Mind std140 layout (pass amat4) and the web shim repacking, then rebuild the shader bundle.Priority: correctness, lower than the winding fix. Many models bake non-uniform scale into geometry or avoid it, so real-world impact is limited.