Lite: Negative scale gltf model is incorrect

s207152 · July 17, 2026, 1:02pm

Related gltf sample:

Pryme8 · July 17, 2026, 5:13pm

maybe @RaananW would have input? Im not sure who did the matrix stuff, but this seems legit.

RaananW · July 20, 2026, 9:16am

Checking now, thanks for reporting

RaananW · July 20, 2026, 2:57pm

github.com/BabylonJS/Babylon-Lite

fix(material): correct double-sided normal flip on negative-scale glTF meshes (#428)

master ← raananw-miniature-doodle

opened 10:25AM - 20 Jul 26 UTC

RaananW

+261 -5

## Summary Fixes the forum report [Lite: negative scale glTF model is incorrect…](https://forum.babylonjs.com/t/lite-negative-scale-gltf-model-is-incorrect/63820). The Khronos **NegativeScaleTest.glb** rendered its mirrored, double-sided reflective spheres **black** in Babylon Lite while Babylon.js rendered them correctly. ## Root cause Mirrored glTF nodes (negative world-matrix determinant, e.g. `scale: [-1,-1,-1]`) have reversed triangle winding relative to the RH→LH root flip. Lite reversed the winding by flipping the pipeline **cull face** (`cullMode: "front"`) while keeping `frontFace: "ccw"`. WebGPU computes `@builtin(front_facing)` from the pipeline's **front-face winding**, *not* the cull face. So for a mirrored mesh, the visible outer surface still evaluated `front_facing == false`. On **double-sided** materials the fragment shader then applies its front-facing normal flip: ```wgsl if (!frontFacing) { N = -N; } ``` which inverted the (already correct) outward normal → `N·V < 0` → the reflective spheres went black. Single-sided meshes were unaffected because they have no such flip — which is why the existing single-sided negative-scale test (scene257) always passed and the bug only showed on double-sided mirrored geometry. ## Fix Reverse the winding by flipping `frontFace` (`ccw`→`cw`) instead of the cull face, matching Babylon.js's `sideOrientation` flip on negative determinant. This keeps `front_facing` consistent with the actual geometry. The non-mirrored path is byte-for-byte unchanged, so all other meshes are unaffected. ```diff - cullMode: noCull || hasDoubleSided ? "none" : reverseWinding ? "front" : "back", - frontFace: "ccw", + cullMode: noCull || hasDoubleSided ? "none" : "back", + frontFace: reverseWinding ? "cw" : "ccw", ``` ## Regression coverage Adds **scene266** — a NegativeScaleTest.glb double-sided sphere grid (each material rendered as an un-mirrored copy and a negative-determinant mirrored copy), with golden-from-BJS parity, a bundle-size ceiling, and a thumbnail. ## Validation - **scene266 parity**: passes (full-image MAD 0.118, every pixel within 5/255 of the BJS reference; sibling IBL scene265 uses maxMad 0.5). - **scene257 parity** (single-sided negative scale): still passes → no regression. - **Full parity suite**: 433 passed. The only failure is scene116, a pre-existing flake already `skipParityOnCI: true` (its own `skipNotes` document a flaky BJS reference capture, not a Lite regression) and provably outside this change's code path. - **Bundle size**: scene266 79.1 KB raw ≤ 79.5 KB ceiling. - **Lint / typecheck**: clean.

Topic		Replies	Views
GL state has incorrect front face when rendering a 2 sided lighting-material on a mesh with a negative determinant world transform Bugs	1	386	November 14, 2023
GLB Export - displaced geometries and flipped orientations Bugs glb	15	280	April 24, 2025
Mirroring mesh without negative scaling Questions	3	557	April 12, 2023
NME PBRMetallicRoughnessBlock reflections are incorrect with right handed coordinate system Bugs	6	448	March 27, 2022
Face ordering problems when replacing material on GLTF-imported mesh Questions	8	1280	March 12, 2019

Lite: Negative scale gltf model is incorrect

Related topics