Repository navigation
Light tree: orientation cone of a reverse-sided light points the wrong way #332
Description
Summary. With adaptive light sampling, reverse-sided lights with a spread below 1 are under-selected: the image is about 4% too dark, and BSDF samples do not compensate. Version 2026.29.1, scalar mode.
Reproduce. Eight DiskLights (radius 0.2, spread 0.5) facing down onto a diffuse floor, max_depth 1, light samples only. Ratio of light_sampling_mode adaptive to uniform:
| lights | adaptive / uniform |
|---|---|
sidedness reverse |
0.9558 |
| the same lights with regular sidedness (rotated instead) | 0.9998 |
reverse, with bsdf_samples on too |
0.9559 |
With spread 0.8 or 1 there is no loss.
Cause. LightTreeCone(const Light*) (lib/rendering/pbr/light/LightTreeUtil.h) takes the cone axis from light->getDirection(0), the local +z, for every light. A light with sidedness reverse emits toward -z, so with a spread below 1 the tree's orientation bound excludes directions the light does emit into, and the importance of the light is underestimated.
Fix (untested). Use -getDirection() as the cone axis for reverse-sided lights (two-sided lights already set mTwoSided).
Activity