Skip to content

Visual Comparisons — Sobol' & MIS

This page shows the concrete, pixel-level effect of RayON's two key variance-reduction techniques side by side. Drag the vertical slider in each pair to reveal how the right-hand image differs from the left-hand one.

How to read the comparisons

Each scene is shown with three slider pairs, isolating one variable at a time:

Pair Left image Right image What you see
Sobol' effect PCG sampler, MIS off Sobol' sampler, MIS off Noise reduction from quasi-random sampling alone
MIS effect Sobol' sampler, MIS off Sobol' sampler, MIS on Noise reduction from Next-Event-Estimation alone
Combined PCG sampler, MIS off Sobol' sampler, MIS on Full improvement — the real-world default vs worst-case baseline

All renders: 1280 × 720 px, offline CUDA mode. The caustics chapel uses 512 SPP; all other scenes use 64 SPP.

How much noise?

At 64 SPP, differences are dramatic — this is deliberately the "hard" regime where both techniques matter most. In production at 2 048+ SPP the images converge and the differences narrow, but the wall-clock time saving (fewer samples needed) is the same.

Scene 1 — OBJ Statue

rough-mirror gold statue · area light · 64 SPP · BVH enabled

A metallic statue illuminated by a single rectangular area light. The rough-mirror surface scatters light in a broad lobe — NEE (MIS) is very effective here because diffuse-like surfaces benefit most from direct light sampling. Sobol' reduces the structured noise visible in the unlit regions of the ground plane.

Sobol' effect (MIS disabled)

PCG Sobol' OBJ Statue — PCG sampler, MIS off OBJ Statue — Sobol' sampler, MIS off

Left: PCG (pseudo-random)  |  Right: Sobol' (quasi-random) — MIS disabled on both. Sobol' reduces high-frequency grain in the shadow regions and on the ground plane.

MIS / NEE effect (Sobol' sampler)

No MIS MIS on OBJ Statue — Sobol', MIS off OBJ Statue — Sobol', MIS on

Left: MIS disabled  |  Right: MIS enabled — Sobol' sampler on both. Next-Event Estimation dramatically reduces noise on the lit portions of the statue and ground.

Combined effect (baseline vs full)

Baseline Full OBJ Statue — PCG, no MIS OBJ Statue — Sobol' + MIS

Left: PCG + no MIS (worst-case baseline)  |  Right: Sobol' + MIS (production default). The improvement is substantial — equivalent to rendering 8–15× more samples with the baseline.

Scene 2 — Caustics Chapel

glass spheres · small bright area light · 512 SPP · indoor scene

The hardest test: a small, very bright area light illuminating through refractive glass spheres in a dark room. Without MIS, most rays never find the light and the image is dominated by fireflies. At 512 SPP the effect of each technique is still clearly visible.

Sobol' effect (MIS disabled)

PCG Sobol' Caustics — PCG, no MIS Caustics — Sobol', no MIS

Left: PCG  |  Right: Sobol' — MIS disabled on both (512 SPP). Sobol' produces more evenly distributed noise; the caustic light patches are less grainy.

MIS / NEE effect (Sobol' sampler)

No MIS MIS on Caustics — Sobol', no MIS Caustics — Sobol', MIS on

Left: No MIS  |  Right: MIS on — Sobol' sampler on both. MIS eliminates most of the fireflies and dramatically brightens the correctly lit diffuse walls. Direct light contributions which were rare without NEE are now sampled explicitly.

Combined effect (baseline vs full)

Baseline Full Caustics — PCG, no MIS Caustics — Sobol' + MIS

Left: PCG + no MIS  |  Right: Sobol' + MIS. This is the most dramatic scene for both techniques combined: the wall illumination and caustic floor patterns are only visible with the full stack enabled.

Scene 3 — Color Bleed Box

Cornell box · red / green / blue walls · area light · 64 SPP

A classic Cornell-box variant with strongly coloured walls. Color bleeding — the indirect bounce of coloured light onto neutral surfaces — is a purely diffuse effect that benefits significantly from both NEE and quasi-random sampling.

Sobol' effect (MIS disabled)

PCG Sobol' Color Bleed — PCG, no MIS Color Bleed — Sobol', no MIS

Left: PCG  |  Right: Sobol' — MIS disabled. Sobol' distributes samples more evenly, giving smoother color gradients on the diffuse walls.

MIS / NEE effect (Sobol' sampler)

No MIS MIS on Color Bleed — Sobol', no MIS Color Bleed — Sobol', MIS on

Left: No MIS  |  Right: MIS on — Sobol' sampler. NEE resolves the direct illumination on the floor and ceiling cleanly. The soft shadow under the spheres and the colour bleeding on the far wall are both sharper.

Combined effect (baseline vs full)

Baseline Full Color Bleed — PCG, no MIS Color Bleed — Sobol' + MIS

Left: PCG + no MIS  |  Right: Sobol' + MIS.

Scene 4 — Default Scene (area-light only)

rough mirrors · glass sphere · Fibonacci-dot sphere · single area light · 64 SPP · ambient off

The standard RayON default scene with ambient lighting disabled so that all illumination comes from a single rectangular area light. The mix of rough-mirror, glass, and diffuse surfaces exercises all code paths: MIS helps diffuse surfaces; Sobol' helps specular highlights.

Sobol' effect (MIS disabled)

PCG Sobol' Default scene — PCG, no MIS Default scene — Sobol', no MIS

Left: PCG  |  Right: Sobol' — MIS disabled. Look at the unlit ground plane and the rough-mirror sphere: Sobol' produces finer, more uniform grain vs the clumpy PCG noise pattern.

MIS / NEE effect (Sobol' sampler)

No MIS MIS on Default scene — Sobol', no MIS Default scene — Sobol', MIS on

Left: No MIS  |  Right: MIS on — Sobol' on both. The shadow boundary under the glass sphere, the lit side of the coloured balls, and the highlight on the rough-mirror sphere all resolve much more cleanly with NEE.

Combined effect (baseline vs full)

Baseline Full Default scene — PCG, no MIS Default scene — Sobol' + MIS

Left: PCG + no MIS (baseline)  |  Right: Sobol' + MIS (production default). At 64 SPP the default configuration is already approaching 256+ SPP quality with the baseline.

Quantitative summary

Technique Dominant benefit Cost
Sobol' sampler Uniform noise, faster convergence everywhere ~0% overhead (same ops, better coverage)
MIS / NEE Massive firefly reduction; direct lighting resolved at very low SPP +1 shadow ray per diffuse bounce (~25–40% more rays)
Sobol' + MIS Best of both: low-discrepancy + direct sampling Combined but non-additive overhead

For the theoretical background, see: